Annual Industry Reportpreviously “The State of Machine Translation”

The State of Translation Automation 2026

AI translation quality depends on the language, the requirements and the workflow around the model. This report compares complete translation solutions—from AI models and MT engines to quality checks, repair and cooperating agents.

Why read the report:

  • See which solutions meet Basic, Standard and Advanced requirements in each language.
  • Compare how prompts, glossaries, quality checks and repair affect quality, cost and speed.
  • Understand where automated assessments are reliable and where the evidence is limited.
Key takeaways Browse chapters

Evaluated AI models and MT engines

57 models from 37 vendors

Company, product and model names and logos are shown for informational purposes only. All trademarks and logos belong to their respective owners. Their use does not imply affiliation or endorsement.

Compare models

Evaluated solutions: a model plus its harness

We evaluate solutions — a model plus its harness: prompts, glossaries and other context, pre- and post-processing. The same engine with a glossary and without one is two solutions. Since we define quality as meeting requirements, we assess the absolute level of quality, and we increase solution complexity only when a simpler solution cannot meet all the requirements. This is how we build solutions for customers, and that is how we approach it in this report.

  1. 1Evaluated

    Off-the-shelf NMT

    A neural engine with default parameters and no glossary.

  2. 2Evaluated

    Adjusted NMT

    With glossaries, tone of voice, tag handling and other parameters supported by an API.

  3. 3Evaluated

    Off-the-shelf LLM

    An LLM with a simple translation prompt.

  4. 4Evaluated

    Adjusted LLM

    LLM with RAG for glossaries and prompt-tuning for tone of voice, tags and other requirements.

  5. 5Out of scope in 2026

    Fine-tuned models

    NMT domain adaptation and LLM fine-tuning in combination with other customization features.

  6. 6Evaluated

    Two-step solution

    Either NMT or LLM step, followed by an LLM-based automatic post-editing (APE) for missing requirements.

  7. 7Evaluated

    Multi-step

    More than two steps in one workflow to meet standard translation requirements: translate, then validate, then repair.

  8. 8Evaluated

    Multi-agent systems

    Agents responsible for different groups of requirements cooperate to meet all of them.

Explore translation workflows

Three levels of translation quality

We evaluate translation for three levels of translation quality, which altogether cover 10 requirements: general translation quality, tone of voice, terminology, gender neutrality, and six formatting requirements: tags, Unicode character integrity, locale conventions, HTML entities, valid JSON, and spacing at tag boundaries.

  1. 1Basic

    General translation quality — fluency, accuracy, completeness, evaluated using the MQM framework.

  2. 2Standard

    Basic + terminology, tone of voice, formatting, tag handling. Preferential requirements that most businesses have.

    In the 2026 benchmark: A glossary, formal tone of voice, HTML and JSON tag handling, Unicode.
  3. 3Advanced

    Standard + customer-specific context — style, idiomaticity, inclusivity, legal, compliance. Preferential requirements that some businesses have.

    In the 2026 benchmark: Gender neutrality.

Explore quality results

Benchmark language directions: English into 100 languages

The benchmark covers translation from English into 100 languages. For 7 of them we found automatic quality assessment tools not reliable enough and excluded them from the results.

Results by language

Key Takeaways for , , , and

Key takeaway 1 · Method

We assess solutions built from 57 AI models and MT engines against three levels of translation quality.

The 2026 State of Translation Automation measures the quality achievable in each target language. We evaluate solutions: a model plus its harness of prompts, glossaries and other context, pre- and post-processing.

The three quality levels define the requirements we check. Basic covers general translation quality. Standard adds terminology, tone of voice and six formatting requirements. Advanced adds gender neutrality.

The benchmark uses four English business texts totaling ~3,000 words and covers 100 target languages. We add complexity when a simpler solution does not meet the requirements.

Our Language engineers run the evaluation using Intento Language Hub. The same team of forward-deployed engineers implements production AI setups for 60+ enterprise customers across hundreds of language pairs.

See more on Method →

38
vendors
57
models
129
model-and-harness configurations
3
quality levels
3,000
source words
100
target languages
Key takeaway 2 · Method

We measure which translation requirements solutions can meet in each language, and what it takes to meet them.

A metric score alone does not show which requirements a translation meets or what still needs improvement. We assess quality against the Basic, Standard and Advanced requirements, together with confidence in the assessment, solution complexity, cost and latency.

See more on Method →

Key takeaway 3 · Method

We shortlist solutions with quality-estimation metrics, then assess them against translation requirements with Intento LQA.

Through the Intento Language Hub API, we obtain COMET-Kiwi and MetricX scores for language assessment and shortlisting. We use these two low-cost quality-estimation metrics to select 10 to 13 solutions per language for further assessment. They estimate accuracy and fluency, but do not check whether the translation meets terminology, tone-of-voice or gender-neutrality requirements.

A higher metric score does not always mean better compliance. In the glossary comparison, the COMET score fell in most cases where the translation used the approved terminology.

We then assess the shortlisted solutions against the 10 requirements grouped into Basic, Standard and Advanced, using Intento LQA. This automatic assessment combines LLM checks with checks in code.

See the quality checks →  ·  See more on Method →

Key takeaway 4 · Method

We assess general translation quality with three LLM judges and combine their findings with a fourth model.

We build the three LLM judges on Claude Opus 5, GPT-5.6 Sol and Gemini 3.1 Pro, using models from three vendors. Each judge receives the source sentence and its translation, without a reference translation, and identifies errors by type and severity using the MQM framework for classifying translation errors. We anonymize the three error lists before combining the findings with Claude Fable 5.

For gender-neutral translations, the LQA agent consults the gender-neutrality agent. We reassess findings on sentences that refer to a person against the relevant gender rule, so required neutral rewording does not count as a translation error.

See more on Method →

Intento LQA assessment of general translation qualitySourcesentencesandtranslationsJudge 1Claude Opus 5Judge 2GPT-5.6 SolJudge 3Gemini 3.1 ProJudgeagreementAt least 90% ofsegmentsCombinefindingsClaude Fable 5Reassess errors incontextConsult thegender-neutralityagentFor gender-neutraltranslationsWeighted errors per1,000 wordsThe judges assess each source sentence and its translation without a reference translation. Sentences that all three judges find error-freeskip the merge. The final score weights errors by severity. Dashed connection: Consultation for gender-neutral translations.Company, product and model names and logos are shown for informational purposes only. All trademarks and logosbelong to their respective owners. Their use does not imply affiliation or endorsement.
Intento LQA assessment of general translation quality
Key takeaway 5 · Method

We use LLM checks for four linguistic requirements and code to check the six formatting requirements.

We assess general translation quality with three LLM judges built on Claude Opus 5, GPT-5.6 Sol and Gemini 3.1 Pro, and combine their findings with Claude Fable 5. For tone of voice, terminology and gender neutrality, we use GPT-5.6 Sol to check each sentence where the requirement applies and Claude Opus 5, from another vendor, to review the findings.

The formatting checks cover tags, Unicode character integrity, locale conventions, HTML entities, valid JSON and spacing at tag boundaries. These checks give the same result for the same translation every time.

See more on Method →

Assessment methods for the 10 translation requirementsGeneral translation qualityTone of voiceTerminologyGender neutralityTagsUnicode character integrityLocale conventionsHTML entitiesValid JSONSpacing at tag boundariesLLM assessmentThree judges; a fourth model combines findingsLLM assessmentOne model checks; another reviewsChecks in codeRepeatable rulesTerminology also uses exact matching; LLM checks resolve inconclusive matches.
Assessment methods for the 10 translation requirements
Key takeaway 6 · Method

We use the same English source texts across target languages so comparisons use the same source material.

The corpus covers four content types: an editorial article on thinking skills, a legal privacy notice for a loyalty program, an institutional mission statement and planning guide for an international education organization, and an operational travel-safety alert.

See more on Method →

4
source texts
2,978
source words
100
target languages
Key takeaway 7 · Method

We accept an Intento LQA result only when the three LLM judges agree on at least 90% of sentences.

For each sentence, we take the most severe error each judge flags (none, minor, major or critical), and we count agreement when the three severity ratings are at most one level apart. Two judges can name the same error differently, so we compare severity only. The judges agree less where translations have more errors.

In 14 of the 96 languages, fewer than 50% of the shortlisted solutions pass this check, so we cannot assess quality there. These are Hausa, Igbo, Javanese, Kinyarwanda, Kurdish (Kurmanji), Malagasy, Scottish Gaelic, Sesotho, Shona, Somali, Wolof, Xhosa, Yoruba and Zulu, and in Igbo, Malagasy, Xhosa and Zulu no solution passes.

See more on Method →

Shortlisted solutions meeting the judge-agreement threshold, by languageLanguages we assessLanguages we cannot assessHalf of the shortlisted solutions020406080100Igbo: 0IgboMalagasy: 0MalagasyZulu: 0ZuluXhosa: 0XhosaSesotho: 8SesothoWolof: 8WolofYoruba: 8YorubaKinyarwanda: 9KinyarwandaShona: 9ShonaSomali: 9SomaliJavanese: 15JavaneseScottish Gaelic: 25Scottish GaelicHausa: 27HausaKurdish (Kurmanji): 33Kurdish (Kurmanji)Cebuano: 50CebuanoBurmese: 55BurmeseYiddish: 58YiddishIcelandic: 62IcelandicFrisian: 69FrisianSwahili: 73SwahiliTatar: 73TatarKyrgyz: 73KyrgyzUzbek: 73UzbekAssamese: 73AssameseSundanese: 75SundaneseKazakh: 75KazakhKhmer: 75KhmerMongolian: 80MongolianPashto: 80PashtoAzerbaijani: 82AzerbaijaniMalay: 82MalayBasque: 83BasqueIrish: 83IrishLithuanian: 83LithuanianSinhala: 83SinhalaJapanese: 85JapaneseGeorgian: 91GeorgianLao: 91LaoOdia (Oriya): 91Odia (Oriya)Armenian: 91ArmenianNepali: 91NepaliPortuguese (European): 91Portuguese (European)Tajik: 92TajikHebrew: 92HebrewGujarati: 92GujaratiThai: 92ThaiDanish: 92DanishChinese Traditional: 92Chinese TraditionalAmharic: 100AmharicKannada: 100KannadaMarathi: 100MarathiWelsh: 100WelshGerman: 100GermanMalayalam: 100MalayalamTelugu: 100TeluguArabic (MSA): 100Arabic (MSA)Belarusian: 100BelarusianSpanish (European): 100Spanish (European)Tagalog: 100TagalogTurkish: 100TurkishHungarian: 100HungarianVietnamese: 100VietnameseEstonian: 100EstonianIndonesian: 100IndonesianMacedonian: 100MacedonianPunjabi: 100PunjabiTamil: 100TamilUkrainian: 100UkrainianChinese Simplified: 100Chinese SimplifiedPolish: 100PolishRomanian: 100RomanianUrdu: 100UrduBengali: 100BengaliCatalan: 100CatalanFarsi: 100FarsiFinnish: 100FinnishGalician: 100GalicianHindi: 100HindiItalian: 100ItalianKorean: 100KoreanLatvian: 100LatvianDutch: 100DutchRussian: 100RussianAlbanian: 100AlbanianSerbian: 100SerbianSwedish: 100SwedishNorwegian: 100NorwegianBulgarian: 100BulgarianCroatian: 100CroatianSlovak: 100SlovakAfrikaans: 100AfrikaansBosnian: 100BosnianCzech: 100CzechSlovenian: 100SlovenianGreek: 100GreekFrench: 100FrenchShare of the shortlisted one-step solutions in each of the 96 measurable languages, in %. For each sentence, the judges agree when the most severe errors they flag (none, minor, major or critical) are atmost one level apart.
Shortlisted solutions meeting the judge-agreement threshold, by language
Key takeaway 8 · Method

We reassess major and critical errors in context, and ~one in three becomes a minor error or no error.

We assess translations with LLM judges one sentence at a time, so the judges count as errors a pronoun or a tense that depends on the previous sentence, content moved across a sentence boundary, or a heading translated as a noun phrase. We send each major and critical error to Claude Fable 5 with the two sentences before and the one after, in the source and in the translation, and ask whether the error still stands. We downgrade most of the changed errors to minor and drop the rest, and we keep nearly every critical error as critical.

Nearly all solutions that reach the limit of three major errors in the four texts only after reassessment use LLMs. Most MT engine translations remain above that limit. Claude Fable 5 is also one of the models we evaluate, and the re-check changes errors in its own translations as often as in those of the other leading LLMs.

See more on Method →

Major and critical errors after reassessment in context, by source textKeptDowngraded to minorDropped020406080100Editorial articleEditorial article: Kept 60 (60.0%)60Editorial article: Downgraded to minor 29 (29.0%)29Editorial article: Dropped 11 (11.0%)11Education guideEducation guide: Kept 71 (71.0%)71Education guide: Downgraded to minor 21 (21.0%)21Education guide: Dropped 8 (8.0%)8Travel alert (JSON)Travel alert (JSON): Kept 78 (78.0%)78Travel alert (JSON): Downgraded to minor 13 (13.0%)13Travel alert (JSON): Dropped 9 (9.0%)9Privacy noticePrivacy notice: Kept 74 (74.0%)74Privacy notice: Downgraded to minor 22 (22.0%)22Privacy notice: Dropped 4 (4.0%)4% of the major and critical errors in the text13,768 major and critical errors found by the LQA judges in the translations of the shortlisted solutionsin 91 languages, re-checked with the two sentences before and the one after, in the source and thetranslation.
Major and critical errors after reassessment in context, by source text
Key takeaway 9 · Method

In our tests, examples in the prompts reduced LLM-judge agreement, so we use prompts without examples.

We assess each translation with three LLM judges, built on Claude Opus 5, GPT-5.6 Sol and Gemini 3.1 Pro. We tested one set of examples for all languages and one set per language on translations into Spanish and Ukrainian from the 2025 edition of the benchmark, which was assessed by human reviewers.

With the set for all languages, the Gemini 3.1 Pro judge flagged an error in 7% of the sentences, against 36% without examples, while the two other judges barely changed, and we saw the same drop when we repeated the test twice on translations from 2026. Of the Ukrainian sentences where human reviewers marked a major or critical error, the three judges agreed on such an error in 6% with the set for all languages, against 46% without examples.

In one of the examples, an English term left untranslated in a Chinese translation was a minor error. With this example in the prompt, the judges also rated an English term left untranslated in a Ukrainian translation as minor, while human reviewers rated it as major. We rely on the agreement of the three judges rather than on human ratings, because in the 2025 human review, reviewers rated severity inconsistently across languages, and sometimes within one language.

See more on Method →

Key takeaway 10 · Method

We found no consistent vendor preference in the LQA judges based on Gemini 3.1 Pro, GPT-5.6 Sol and Claude Opus 5.

Intento LQA scores translations with a panel of three LLM judges: Gemini 3.1 Pro, GPT-5.6 Sol and Claude Opus 5. Each judge is itself a solution, a model in a harness that scores translations.

Two of the three judge models also appear as translation solutions in the shortlist, so we tested whether a judge scores translations from its own vendor higher. For every scored solution we compared each judge with the average of the three, separately for solutions from the same vendor as the judge and for all others. The same-vendor difference is below 0.2 points on the 0–100 general translation quality score and goes in different directions for different judges, so we found no consistent vendor preference in this test.

The same test showed that the judges differ in strictness. Claude Opus 5, the strictest judge, scores 2.7 points below Gemini 3.1 Pro, the most lenient one, while the best and worst shortlisted solutions in a language differ by about 12 points. A judge that is strict or lenient across the board does not change the order of solutions within a language, because every judge scores every solution. Absolute scores do depend on which judges sit on the panel, so we keep the same panel for the whole 2026 edition.

See more on Method →

Score difference for each judge relative to the other two-2-1012Gemini 3.1 Pro: 1.3+1.30Gemini 3.1 ProGPT-5.6 Sol: 0.13+0.13GPT-5.6 SolClaude Opus 5: -1.44-1.44Claude Opus 5Points on the 0–100 general translation quality score, 293 solutions scored by allthree judges. The best and worst shortlisted solutions in a language differ by about12 points.Company, product and model names and logos are shown for informationalpurposes only. All trademarks and logos belong to their respective owners.Their use does not imply affiliation or endorsement.
Score difference for each judge relative to the other two
Key takeaway 11 · Models

The best open-weight LLM translation has more errors than the best other LLM translation in all 50 languages with results for both.

The median ratio is ~2.9 times as many errors in the best open-weight LLM translation.

The gap is smallest in Indonesian, where Google Gemma 4 has the only open-weight solution that meets general translation quality (no critical error, at most three major errors and at most 20 weighted errors per 1,000 words), and it meets every other Standard requirement too. The open-weight LLMs on our shortlists are Google Gemma 4, OpenAI gpt-oss-120b, Meta Llama 4 Maverick, Alibaba Qwen3-235B-A22B and Swiss AI Apertus-8B.

See details →  ·  See more on Models →

Errors in the best open-weight and other LLM solutions, by languageLanguageSame number of errorsLimit for general translation quality55101020205050100100250250Best open-weight LLM, weighted errors per 1,000 wordsBest other LLM, weighted errors per 1,000 wordsIndonesian: Google Gemma 4 17.8, Google Gemini 3.1 Pro Preview 12.5 weighted errors per 1,000 wordsSwahili: Google Gemma 4 54.7, Anthropic Claude Fable 5 36.3 weighted errors per 1,000 wordsTagalog: Google Gemma 4 35.9, OpenAI GPT-5.6 Terra 21.4 weighted errors per 1,000 wordsCebuano: Google Gemma 4 61.3, Google Gemini 3.1 Pro Preview 34.9 weighted errors per 1,000 wordsKorean: Google Gemma 4 40.5, OpenAI GPT-5.6 Terra 21.4 weighted errors per 1,000 wordsRussian: Google Gemma 4 24.7, Anthropic Claude Fable 5 12.9 weighted errors per 1,000 wordsSerbian: Google Gemma 4 37.2, Anthropic Claude Fable 5 17.8 weighted errors per 1,000 wordsAfrikaans: Google Gemma 4 63.9, OpenAI GPT-5.5 28.7 weighted errors per 1,000 wordsHindi: Google Gemma 4 26.4, Anthropic Claude Fable 5 11.5 weighted errors per 1,000 wordsTurkish: Google Gemma 4 28.3, Anthropic Claude Fable 5 12.2 weighted errors per 1,000 wordsDutch: OpenAI gpt-oss-120b 33.9, OpenAI GPT-5.5 14.5 weighted errors per 1,000 wordsGujarati: Google Gemma 4 45.2, Anthropic Claude Fable 5 18.8 weighted errors per 1,000 wordsChinese Simplified: OpenAI gpt-oss-120b 54.4, OpenAI GPT-5.6 Terra 22.7 weighted errors per 1,000 wordsFrench: OpenAI gpt-oss-120b 54.1, Anthropic Claude Fable 5 22.1 weighted errors per 1,000 wordsFarsi: Google Gemma 4 32, Google Gemini 3.1 Pro Preview 12.9 weighted errors per 1,000 wordsUrdu: Google Gemma 4 45.8, Anthropic Claude Fable 5 18.5 weighted errors per 1,000 wordsKannada: Google Gemma 4 59.3, OpenAI GPT-5.6 Terra 23.7 weighted errors per 1,000 wordsThai: Google Gemma 4 49.8, Anthropic Claude Opus 4.8 19.8 weighted errors per 1,000 wordsNorwegian: Google Gemma 4 25, OpenAI GPT-5.5 9.9 weighted errors per 1,000 wordsItalian: Google Gemma 4 28.7, OpenAI GPT-5.5 11.2 weighted errors per 1,000 wordsVietnamese: Google Gemma 4 35.3, Anthropic Claude Opus 4.8 13.8 weighted errors per 1,000 wordsBosnian: Google Gemma 4 46.1, Google Gemini 3.1 Pro Preview 17.5 weighted errors per 1,000 wordsArabic (MSA): Google Gemma 4 43.2, OpenAI GPT-5.5 16.2 weighted errors per 1,000 wordsSundanese: OpenAI gpt-oss-120b 97.9, OpenAI GPT-5.5 35.3 weighted errors per 1,000 wordsAlbanian: Google Gemma 4 51.7, Anthropic Claude Fable 5 18.1 weighted errors per 1,000 wordsUkrainian: Google Gemma 4 28.3, Anthropic Claude Fable 5 9.9 weighted errors per 1,000 wordsKhmer: Google Gemma 4 53.4, OpenAI GPT-5.6 Sol 18.5 weighted errors per 1,000 wordsMalayalam: Google Gemma 4 51.7, OpenAI GPT-5.6 Sol 16.8 weighted errors per 1,000 wordsPortuguese (European): Meta Llama 4 Maverick 46.1, Anthropic Claude Fable 5 14.8 weighted errors per 1,000 wordsAssamese: Google Gemma 4 95.6, OpenAI GPT-5.5 30.3 weighted errors per 1,000 wordsPolish: Google Gemma 4 43.5, OpenAI GPT-5.6 Terra 13.5 weighted errors per 1,000 wordsMarathi: Google Gemma 4 54.4, OpenAI GPT-5.6 Sol 16.2 weighted errors per 1,000 wordsAzerbaijani: Google Gemma 4 56, OpenAI GPT-5.5 16.5 weighted errors per 1,000 wordsOdia (Oriya): Google Gemma 4 79.1, OpenAI GPT-5.6 Luna 23.1 weighted errors per 1,000 wordsBengali: Google Gemma 4 59, OpenAI GPT-5.5 17.1 weighted errors per 1,000 wordsGerman: Meta Llama 4 Maverick 36.9, Google Gemini 3.1 Pro Preview 10.5 weighted errors per 1,000 wordsSinhala: Google Gemma 4 60, Anthropic Claude Fable 5 17.1 weighted errors per 1,000 wordsMacedonian: Google Gemma 4 42.8, OpenAI GPT-5.6 Terra 11.9 weighted errors per 1,000 wordsCatalan: OpenAI gpt-oss-120b 46.8, Google Gemini 3.1 Pro Preview 11.9 weighted errors per 1,000 wordsMalay: OpenAI gpt-oss-120b 43.5, OpenAI GPT-5.5 10.9 weighted errors per 1,000 wordsPunjabi: Google Gemma 4 69.9, OpenAI GPT-5.6 Sol 16.8 weighted errors per 1,000 wordsAmharic: Google Gemma 4 104.5, OpenAI GPT-5.6 Sol 23.1 weighted errors per 1,000 wordsHebrew: OpenAI gpt-oss-120b 75.8, Anthropic Claude Fable 5 16.5 weighted errors per 1,000 wordsTelugu: Google Gemma 4 67.6, Anthropic Claude Fable 5 14.5 weighted errors per 1,000 wordsCroatian: Google Gemma 4 47.8, OpenAI GPT-5.5 9.6 weighted errors per 1,000 wordsBulgarian: OpenAI gpt-oss-120b 63.3, Google Gemini 3.1 Pro Preview 11.2 weighted errors per 1,000 wordsTamil: Google Gemma 4 65.6, OpenAI GPT-5.6 Sol 10.5 weighted errors per 1,000 wordsTatar: OpenAI gpt-oss-120b 206, Google Gemini 3.1 Pro Preview 30 weighted errors per 1,000 wordsPashto: OpenAI gpt-oss-120b 182.6, OpenAI GPT-5.6 Sol 26.4 weighted errors per 1,000 wordsWelsh: OpenAI gpt-oss-120b 189.5, Google Gemini 3.1 Pro Preview 12.9 weighted errors per 1,000 wordsIndonesianSwahiliWelshOne point per language, 50 languages, log scales. Below the dashed line, the best open-weight solution has moreerrors than the best other LLM solution. Weighted errors: general translation quality errors per 1,000 words,weighted by severity (minor 1, major 5, critical 10).Company, product and model names and logos are shown for informational purposes only. Alltrademarks and logos belong to their respective owners. Their use does not imply affiliation orendorsement.
Errors in the best open-weight and other LLM solutions, by language
Key takeaway 12 · Models

Some translation errors and unmet requirements recur across languages for the same model.

The translated JSON document is invalid in every language for Translated Lara 2, Microsoft Translator, Google Translation LLM and Widn AI Tower Sugarloaf. Translations of the JSON document from Anthropic Claude Opus 4.8 meet the tag handling requirement in 12 of 46 languages, against 73 of 76 languages for Google Gemini 3.1 Pro Preview in every document.

Translations from Translated Lara 2 carry an HTML code in place of each newline character between the tags and the JSON fields of the source file, in every language. Without these codes, the translated JSON document would be valid in 36 of the 44 languages. In nearly every language, translations from Google Translation LLM carry additions (sentences that are not in the source).

We translate every text with its HTML tags and set the format parameter for tagged content where the API has one, textType html for Microsoft Translator, mimeType text/html for Google Translation LLM and content_type text/html for Translated Lara 2. We don't set a format parameter for Widn AI Tower Sugarloaf.

See details →  ·  See more on Models →

Share of one-step solutions meeting the tag handling and formatting requirements, by modelTag handlingFormatting0%20%40%60%80%100%Google Gemini 3.1 Pro Preview, Tag handling: 96%96%Google Gemini 3.1 Pro Preview, Formatting: 87%87%Google Gemini 3.1 Pro PreviewGoogle Gemini 3.5 Flash, Tag handling: 56%56%Google Gemini 3.5 Flash, Formatting: 90%90%Google Gemini 3.5 FlashOpenAI GPT-5.5, Tag handling: 89%89%OpenAI GPT-5.5, Formatting: 76%76%OpenAI GPT-5.5Anthropic Claude Fable 5, Tag handling: 68%68%Anthropic Claude Fable 5, Formatting: 95%95%Anthropic Claude Fable 5DeepL Next Gen (off the shelf), Tag handling: 0%0%DeepL Next Gen (off the shelf), Formatting: 0%0%DeepL Next Gen (off the shelf)Anthropic Claude Opus 4.8, Tag handling: 24%24%Anthropic Claude Opus 4.8, Formatting: 91%91%Anthropic Claude Opus 4.8OpenAI GPT-4.1, Tag handling: 71%71%OpenAI GPT-4.1, Formatting: 98%98%OpenAI GPT-4.1Translated Lara 2, Tag handling: 70%70%Translated Lara 2, Formatting: 0%0%Translated Lara 2Google Gemma 4, Tag handling: 30%30%Google Gemma 4, Formatting: 78%78%Google Gemma 4Anthropic Claude Sonnet 5, Tag handling: 31%31%Anthropic Claude Sonnet 5, Formatting: 86%86%Anthropic Claude Sonnet 5GLM-5, Tag handling: 88%88%GLM-5, Formatting: 91%91%GLM-5OpenAI GPT-5.6 Sol, Tag handling: 61%61%OpenAI GPT-5.6 Sol, Formatting: 97%97%OpenAI GPT-5.6 SolMicrosoft Translator (off the shelf), Tag handling: 0%0%Microsoft Translator (off the shelf), Formatting: 0%0%Microsoft Translator (off the shelf)Anthropic Claude Sonnet 4.6, Tag handling: 19%19%Anthropic Claude Sonnet 4.6, Formatting: 68%68%Anthropic Claude Sonnet 4.6Google Translation LLM (adjusted), Tag handling: 0%0%Google Translation LLM (adjusted), Formatting: 0%0%Google Translation LLM (adjusted)Mistral Large 3, Tag handling: 36%36%Mistral Large 3, Formatting: 43%43%Mistral Large 3OpenAI GPT-5.6 Luna, Tag handling: 46%46%OpenAI GPT-5.6 Luna, Formatting: 85%85%OpenAI GPT-5.6 LunaCohere Command A (command-a-translate-08-2025), Tag handling: 27%27%Cohere Command A (command-a-translate-08-2025), Formatting: 91%91%Cohere Command A (command-a-translate-08-2025)OpenAI GPT-5.6 Terra, Tag handling: 80%80%OpenAI GPT-5.6 Terra, Formatting: 85%85%OpenAI GPT-5.6 TerraGrok 4.20 (reasoning), Tag handling: 38%38%Grok 4.20 (reasoning), Formatting: 88%88%Grok 4.20 (reasoning)DeepL Next Gen (adjusted), Tag handling: 0%0%DeepL Next Gen (adjusted), Formatting: 13%13%DeepL Next Gen (adjusted)DeepSeek V3.2, Tag handling: 57%57%DeepSeek V3.2, Formatting: 93%93%DeepSeek V3.2MoonshotAI Kimi K2.5, Tag handling: 42%42%MoonshotAI Kimi K2.5, Formatting: 25%25%MoonshotAI Kimi K2.5OpenAI gpt-oss-120b, Tag handling: 42%42%OpenAI gpt-oss-120b, Formatting: 83%83%OpenAI gpt-oss-120bModels with 12 or more languages. Off the shelf and adjusted: MT engines without and with a glossary and a formality setting. The other models are LLMs adjusted with tone of voice,terminology and other context in the request.Company, product and model names and logos are shown for informational purposes only. All trademarks and logos belong to their respective owners. Their usedoes not imply affiliation or endorsement.
Share of one-step solutions meeting the tag handling and formatting requirements, by model
Key takeaway 13 · Models

We verify language support by checking whether each model returns a usable translation in that language.

Vendor documentation does not reliably say which languages a model supports. LLMs have no language list, and when we tested the language lists of MT vendors, they were wrong in both directions: we did not get a translation in some listed languages, and got translations in some unlisted ones.

From 21 of the 57 models, in 94 of the 100 languages, we got text in the wrong language, garbage or a near-empty output, a translation cut off part-way, text in the wrong script, or text far shorter than the source. We remove such outputs before scoring quality, because a quality score for a text that is not a translation means nothing. A model supports a language only when we got a translation in it from the model, and that is how we determine language support for every model in the report.

See details →  ·  Explore interactively →  ·  See more on Models →

Language support by modelSupported languageNominally supported language020406080100Anthropic Claude Fable 5: 100Anthropic Claude Fable 5Anthropic Claude Opus 4.8: 100Anthropic Claude Opus 4.8Anthropic Claude Sonnet 4.6: 100Anthropic Claude Sonnet 4.6Anthropic Claude Sonnet 5: 100Anthropic Claude Sonnet 5DeepSeek V3.2: 100DeepSeek V3.2GLM-5: 100GLM-5Google Gemini 3.1 Pro Preview: 100Google Gemini 3.1 Pro PreviewGoogle Gemini 3.5 Flash: 100Google Gemini 3.5 FlashGoogle Gemma 4: 100Google Gemma 4Grok 4.20 (non-reasoning): 100Grok 4.20 (non-reasoning)Grok 4.20 (reasoning): 100Grok 4.20 (reasoning)Mistral Large 3: 100Mistral Large 3Mistral Medium 3.1: 100Mistral Medium 3.1MoonshotAI Kimi K2.5: 100MoonshotAI Kimi K2.5OpenAI GPT-4.1: 100OpenAI GPT-4.1OpenAI GPT-5.5: 100OpenAI GPT-5.5OpenAI GPT-5.6 Luna: 100OpenAI GPT-5.6 LunaOpenAI GPT-5.6 Sol: 100OpenAI GPT-5.6 SolOpenAI GPT-5.6 Terra: 100OpenAI GPT-5.6 TerraAnthropic Claude Haiku 4.5: 99Anthropic Claude Haiku 4.5: 1 Nominally supported languageAnthropic Claude Haiku 4.5Google Cloud Advanced Translation API: 98Google Cloud Advanced Translation API: 1 Nominally supported languageGoogle Cloud Advanced Translation APIM2.5: 98M2.5: 2 Nominally supported languageM2.5OpenAI gpt-oss-120b: 98OpenAI gpt-oss-120b: 2 Nominally supported languageOpenAI gpt-oss-120bTranslated Lara 2: 97Translated Lara 2Cohere Command A (command-a-translate-08-2025): 95Cohere Command A (command-a-translate-08-2025): 5 Nominally supported languageCohere Command A (command-a-translate-08-2025)ModernMT Enterprise Edition: 93ModernMT Enterprise EditionMicrosoft Translator: 89Microsoft Translator: 1 Nominally supported languageMicrosoft TranslatorDeepL Next Gen: 84DeepL Next Gen: 1 Nominally supported languageDeepL Next GenApertus-8B-2509: 83Apertus-8B-2509: 17 Nominally supported languageApertus-8B-2509Yandex: 83YandexAlibaba Cloud Qwen3-235B-A22B: 79Alibaba Cloud Qwen3-235B-A22BGoogle Translation LLM: 79Google Translation LLM: 1 Nominally supported languageGoogle Translation LLMAlibaba: 77Alibaba: 1 Nominally supported languageAlibabaAmazon: 69AmazonCohere Tiny Aya: 68Cohere Tiny Aya: 1 Nominally supported languageCohere Tiny AyaAI21 Jamba Large 1.7: 65AI21 Jamba Large 1.7: 35 Nominally supported languageAI21 Jamba Large 1.7SYSTRAN PNMT API: 57SYSTRAN PNMT APIAppTek NMT: 54AppTek NMTTilde: 28Tilde: 3 Nominally supported languageTildeSupertext Supertext LLM: 26Supertext Supertext LLMNiuTrans: 22NiuTransPangeanic: 22PangeanicWidn AI Tower Sugarloaf: 22Widn AI Tower SugarloafOracle: 21OracleTencent Cloud TMT: 13Tencent Cloud TMTBaidu: 12BaiduMeta Llama 4 Maverick: 11Meta Llama 4 MaverickNaver Papago NMT Commercial: 11Naver Papago NMT CommercialSpeakLeash Bielik-11B-v3.0-Instruct: 11SpeakLeash Bielik-11B-v3.0-Instruct: 19 Nominally supported languageSpeakLeash Bielik-11B-v3.0-InstructWidn AI Tower Anthill: 11Widn AI Tower Anthill: 27 Nominally supported languageWidn AI Tower AnthillYoudao: 8Youdao: 2 Nominally supported languageYoudaoTartuNLP: 7TartuNLPALL-Lab: 6ALL-Lab: 8 Nominally supported languageALL-LabiFlyTek MT: 6iFlyTek MTElia Elhuyar Machine Translation System: 5Elia Elhuyar Machine Translation SystemWidn AI Tower Vesuvius: 2Widn AI Tower Vesuvius: 36 Nominally supported languageWidn AI Tower VesuviusTarjama: 1TarjamaOut of 100 languages. Supported: we got a translation from at least one solution of the model. Nominally supported: we got output in the language from the model, but no outputpassed all five checks. Without the languages outside the license of one MT engine.Company, product and model names and logos are shown for informational purposes only. All trademarks and logos belong to their respective owners. Theiruse does not imply affiliation or endorsement.
Language support by model
Key takeaway 14 · Models

Vendor rules for access, billing, region and request size can affect translation cost.

The rules fall into four kinds: access (keys, trials, licenses), billing and quota (prepaid balances, character quotas, minimum charges), region (endpoints and resources bound to one region) and request limits (parameters and request sizes). Some vendors have a minimum charge per request (5,000-20,000 characters), which raises the cost of short texts. During this benchmark we use Intento Language Hub, which provides a single API to all NMT and LLM models and handles differences in how the APIs express translation requests.

See more on Models →

Key takeaway 15 · Models

We use zero data retention wherever the model API supports it.

Zero data retention means the vendor does not store the requests sent to its API, under the applicable retention terms. For enterprise translation this matters because the content is often confidential, and the texts in this benchmark are enterprise content, contributed on our promise to use them for evaluation only, never for training.

The API of Claude Fable 5 returns an error unless the account has data retention turned on, so we translate with it in a dedicated workspace. The workspace has its own API key and keeps requests for 30 days, and we use it only for the benchmark texts.

See more on Models →

Key takeaway 16 · Harnesses

General translation quality improves mainly in three-step solutions (translation + LQA + repair), while most translations meet the other Standard requirements in two-step solutions (translation + post-edit) already.

In a two-step solution, we post-edit the one-step translation with an LLM, using a prompt for the requirement it falls short on, and repair tags and formatting by code. In a three-step solution, we assess the translation with Intento LQA and send it with the findings to an LLM for a repair. In most solutions, we post-edit or repair with the model we translated the text with. We implement these two-step and three-step workflows in Intento Language Hub, including the transfer of Intento LQA findings to the repair step.

On the same solutions, weighted errors fall from 36 per 1,000 words in the one-step solutions to 28 in the two-step solutions and to 16 in the three-step solutions. General translation quality comes mostly from the pre-training of a model, and we see little gain from a second pass without a list of the errors.

Where only tone of voice or terminology fell short, 90% of the two-step solutions met every requirement, and where only tags or formatting fell short, ~80% did, with tags and formatting repaired by code. Three-step solutions meet these requirements about as often. We assess every two-step and three-step solution again, without the findings in the prompts of the LQA judges.

See more on Harnesses →

Requirements missed per 1,000 words in one-step, two-step and three-step solutionsOne-step solutionsTwo-step solutions (translation + post-edit)Three-step solutions (translation + LQA + repair)010203040General translation quality, One-step solutions: 36.136.1General translation quality, Two-step solutions (translation + post-edit): 27.527.5General translation quality, Three-step solutions (translation + LQA + repair): 15.815.8General translation qualityTone of voice, One-step solutions: 0.70.7Tone of voice, Two-step solutions (translation + post-edit): 0.30.3Tone of voice, Three-step solutions (translation + LQA + repair): 00Tone of voiceTag handling, One-step solutions: 1.31.3Tag handling, Two-step solutions (translation + post-edit): 0.70.7Tag handling, Three-step solutions (translation + LQA + repair): 00Tag handlingFormatting, One-step solutions: 0.30.3Formatting, Two-step solutions (translation + post-edit): 00Formatting, Three-step solutions (translation + LQA + repair): 00FormattingMedian over the same solutions, among the one-step solutions that missed the requirement. General translation quality: errorsweighted by severity (minor 1, major 5, critical 10). Tone of voice: sentences in the wrong register. Tag handling: missing or extratags. Formatting: failed checks. Terminology is left out, with too few solutions.
Requirements missed per 1,000 words in one-step, two-step and three-step solutions
Key takeaway 17 · Harnesses

~82% of the three-step solutions (translation + LQA + repair) still meet every requirement met by the one-step solution, and the others most often miss terminology or tone of voice.

A solution meets general translation quality with no critical error, at most three major errors and at most 20 weighted errors per 1,000 words. Of the three-step solutions that meet it but not every other requirement, 25% still meet every requirement met by the one-step solution.

In the languages with a glossary, 58% of the three-step solutions meet the terminology requirement, against 88% of the one-step solutions they are built on. Most of the terms missing in the three-step solutions are paraphrased. ~5% of the three-step solutions have a critical error the one-step solution did not have. We check every requirement again in every three-step solution.

See details →  ·  See more on Harnesses →

Key takeaway 18 · Harnesses

Gender-neutral versions of the best solutions lose less accuracy with the gender-neutrality agent involved at all stages of the solution, including post-editing, LQA and repair, than with the agent involved in the translation step only.

Rewording for gender neutrality changes the meaning of some sentences, and Intento LQA counts those changes as errors. With the gender-neutrality agent involved in the translation step only, 39% of the gender-neutral versions still met every Standard requirement. With the agent involved at all stages of the solution, including post-editing, LQA and repair, 85% did.

Of the gender-neutral versions that still fell short, 16% met the Advanced requirements as two-step solutions (translation + post-edit), and 46% of the rest as three-step solutions (translation + LQA + repair), with the gender findings.

We use Intento Language Hub to coordinate the gender-neutrality agent with translation, post-editing, LQA and repair, using the same language-specific gender rule throughout.

See more on Harnesses →

Gender-neutral versions still meeting every Standard requirementStill meet every Standard requirementFall short020406080100Agent in the translation step onlyAgent in the translation step only: Still meet every Standard requirement 39 (39.0%)Agent in the translation step only: Fall short 61 (61.0%)39%Agent at all stages, including post-editing, LQA and repairAgent at all stages, including post-editing, LQA and repair: Still meet every Standard requirement 85 (85.0%)Agent at all stages, including post-editing, LQA and repair: Fall short 15 (15.0%)85%% of the gender-neutral versionsPercent of the first gender-neutral versions of solutions that met everyStandard requirement before.
Gender-neutral versions still meeting every Standard requirement
Key takeaway 19 · Harnesses

13 of the 28 MT engines support a glossary and 8 support tone of voice through their API, and 14 support neither.

An MT engine takes the Standard requirements only through its API, if the model supports them. We upload a glossary in advance and reference it in each request, send the terms found in each segment with the request, or put those terms into a free-text instructions field.

Tone values differ by engine: formal, natural, expressive or precise at ALL-Lab, politeness as more or less at Supertext, formality as more, less, prefer_more or prefer_less at DeepL Next Gen. Naver accepts tone only in Korean and Japanese, DeepL Next Gen in 13 languages and Supertext in 26.

See details →  ·  See more on Harnesses →

Key takeaway 20 · Harnesses

We run a dedicated test for each of three types of glossary support, because advertised support is not always present, often comes with limits, and can fail without an error.

"Supports glossaries" means one of three things across the models we tested: terms inserted into the prompt, a glossary uploaded in advance, or a dynamic glossary sent with each request. For a model with glossary support in a language, we benchmark two translations: a baseline, and one adjusted with the glossary and the other Standard requirements. For a model without glossary support in a language, we stop at the baseline. In both cases, we then try to improve the best results further by adding workflow steps.

See details →  ·  See more on Harnesses →

Key takeaway 21 · Harnesses

Sending a glossary to an MT engine does not guarantee its use, so we test it with texts that must change when the glossary applies.

MT engines accept a glossary through the API and report success. Often the translation does not change, and the API response does not show it. So we built a test set that contains the glossary terms. We translate it with the glossary and without it, and count the approved terms in each result. In 2026 the test set has 46 terms in 16 languages.

It caught three mistakes on our side that any integration can make. A glossary created for Google Cloud Advanced Translation does not apply to Google Translation LLM, although both sit behind one API. ModernMT imports a glossary in the background, and the id it returns is empty until the import completes. Yandex takes the glossary only inside each request, not as an uploaded resource.

The test set also showed where engines do not apply a glossary at all: Yandex rejects it for Belarusian, Ukrainian and Serbian, Google Translation LLM rejects it for Serbian in Latin script, and ModernMT ignores it for Arabic.

See more on Harnesses →

Key takeaway 22 · Harnesses

We use one prompt template for all 100 languages, with three parameters (the target language and its script, the tone instruction and the glossary terms in the segment).

There are two templates: a short one for the Basic quality level and a longer one for the Standard level. We name the script where models diverge: Latin for Serbian and Bosnian, Cyrillic for Kazakh, Hebrew script for Yiddish, Kurmanji for Kurdish. The tone instruction stays the same across a language. We match the glossary terms per segment. NMT models do not support prompts, so we pass the glossary and the tone of voice through API parameters, if the model supports them.

For requests with context, we use Intento Language Hub to retrieve the relevant glossary terms and language-specific tone instructions for the prompt.

See more on Harnesses →

Key takeaway 23 · Languages

We measure the cost, time and complexity needed for two solutions per language to meet each set of quality requirements.

We seek a main solution and a backup. Two solutions meet the requirements in 79 of the 100 languages for Basic, in 78 for Standard and in 53 for Advanced, and in 18 languages we cannot assess all the requirements. For each language and level, we also report the errors per 1,000 words of the best solution.

We implement the candidate solutions in Intento Language Hub and compare the quality, cost and time needed to meet the requirements in each language.

See details →  ·  Explore interactively →  ·  See more on Languages →

Languages by the number of solutions meeting each set of requirementsTwo or more solutionsOne solutionNo solutionCannot assess020406080100BasicBasic: Two or more solutions 79 (79.0%)79Basic: One solution 1 (1.0%)Basic: No solution 2 (2.0%)Basic: Cannot assess 18 (18.0%)18StandardStandard: Two or more solutions 78 (78.0%)78Standard: One solution 2 (2.0%)Standard: No solution 2 (2.0%)Standard: Cannot assess 18 (18.0%)18AdvancedAdvanced: Two or more solutions 53 (53.0%)53Advanced: One solution 18 (18.0%)18Advanced: No solution 11 (11.0%)11Advanced: Cannot assess 18 (18.0%)18languages100 languages. A solution meets a level when it meets all its requirements. Advanced counts the two best solutions of eachlanguage. Cannot assess: we cannot assess all the requirements in the language.
Languages by the number of solutions meeting each set of requirements
Quality levelBasicStandardAdvancedCustom
Two or more solutionsOne solutionNo solutionCannot assess

On each language: ★ for each solution that meets the selected requirements, the total errors per 1,000 words (EPT) of the best of them across all requirements, the seconds per 1,000 words of the fastest and the cost per million characters of the cheapest. Where no solution meets the selected requirements, the values of the best available solution.

Amharic★★8.9 EPT321 sec$1,043/1M
Armenian★★7.3 EPT254 sec$556/1M
Assamese★★9.9 EPT292 sec$979/1M
Azerbaijani★★15.2 EPT74 sec$7.78/1M
Basque★★7.9 EPT209 sec$574/1M
Belarusian★★6.6 EPT181 sec$77/1M
Bengali★★6.6 EPT144 sec$12/1M
Bulgarian★★7.3 EPT116 sec$4.94/1M
Catalan★★9.2 EPT68 sec$4.52/1M
Cebuano★★11.2 EPT293 sec$604/1M
Chinese Simplified★★13.2 EPT238 sec$589/1M
Chinese Traditional★★11.2 EPT172 sec$490/1M
Croatian★★4.6 EPT172 sec$45/1M
Danish★★12.2 EPT25 sec$4.09/1M
Dutch★★2.6 EPT56 sec$9.66/1M
Estonian★★5.6 EPT202 sec$607/1M
Farsi★★12.5 EPT68 sec$4.41/1M
Finnish★★13.2 EPT67 sec$7.60/1M
Galician★★10.5 EPT35 sec$7.20/1M
Georgian★★16.2 EPT120 sec$12/1M
German★★8.6 EPT94 sec$4.02/1M
Greek★★2.6 EPT249 sec$47/1M
Hungarian★★12.5 EPT147 sec$33/1M
Icelandic★★10.5 EPT237 sec$47/1M
Indonesian★★12.5 EPT12 sec$0.38/1M
Irish★★8.6 EPT141 sec$951/1M
Japanese★★10.9 EPT191 sec$645/1M
Kannada★★6.3 EPT280 sec$948/1M
Kazakh★★8.9 EPT108 sec$11/1M
Khmer★★7.3 EPT432 sec$27/1M
Korean★★9.9 EPT293 sec$577/1M
Kyrgyz★★7.6 EPT196 sec$553/1M
Lao★★15.2 EPT163 sec$28/1M
Malay★★9.6 EPT13 sec$0.40/1M
Malayalam★★9.9 EPT213 sec$28/1M
Marathi★★5.9 EPT277 sec$902/1M
Mongolian★★9.9 EPT325 sec$609/1M
Norwegian★★9.9 EPT50 sec$10/1M
Odia (Oriya)★★11.5 EPT360 sec$1,054/1M
Slovak★★4.0 EPT108 sec$4.90/1M
Swahili★★8.2 EPT119 sec$543/1M
Swedish★★13.8 EPT41 sec$4.23/1M
Tagalog★★7.6 EPT216 sec$566/1M
Tajik★★7.3 EPT271 sec$577/1M
Tamil★★10.5 EPT200 sec$9.10/1M
Tatar★★7.9 EPT255 sec$645/1M
Telugu★★5.6 EPT273 sec$929/1M
Thai★★9.2 EPT180 sec$554/1M
Turkish★★10.9 EPT45 sec$10/1M
Ukrainian★★5.9 EPT94 sec$20/1M
Uzbek★★5.6 EPT223 sec$38/1M
Vietnamese★★12.5 EPT67 sec$17/1M
Welsh★★8.9 EPT108 sec$4.87/1M
Afrikaans★18.1 EPT203 sec$968/1M
Bosnian★6.3 EPT267 sec$993/1M
Frisian★14.2 EPT301 sec$964/1M
Gujarati★15.2 EPT200 sec$29/1M
Hindi★4.3 EPT228 sec$1,045/1M
Italian★12.2 EPT118 sec$38/1M
Latvian★12.5 EPT274 sec$972/1M
Lithuanian★11.9 EPT359 sec$1,051/1M
Macedonian★12.2 EPT64 sec$13/1M
Nepali★17.1 EPT118 sec$11/1M
Pashto★13.2 EPT403 sec$1,071/1M
Portuguese (European)★14.2 EPT104 sec$72/1M
Punjabi★14.8 EPT679 sec$1,000/1M
Romanian★4.0 EPT216 sec$1,029/1M
Russian★14.2 EPT201 sec$29/1M
Serbian★9.2 EPT225 sec$1,034/1M
Slovenian★14.2 EPT192 sec$49/1M
Spanish (European)★6.3 EPT24 sec$3.92/1M
Albanian17.2 EPT158 sec$63/1M
Arabic (MSA)8.9 EPT239 sec$1,047/1M
Burmese11.5 EPT730 sec$1,132/1M
Czech6.2 EPT232 sec$1,032/1M
French17.2 EPT192 sec$995/1M
Hebrew18.2 EPT300 sec$919/1M
Polish7.1 EPT197 sec$900/1M
Sinhala7.9 EPT147 sec$926/1M
Sundanese21.4 EPT243 sec$636/1M
Urdu5.0 EPT303 sec$991/1M
Yiddish10.5 EPT440 sec$1,116/1M
Haitian Creolecannot assess
Hausacannot assess
Igbocannot assess
Javanesecannot assess
Kinyarwandacannot assess
Kurdish (Kurmanji)cannot assess
Malagasycannot assess
Maltesecannot assess
Scottish Gaeliccannot assess
Sesothocannot assess
Shonacannot assess
Somalicannot assess
Turkmencannot assess
Uyghurcannot assess
Wolofcannot assess
Xhosacannot assess
Yorubacannot assess
Zulucannot assess
+ 90 more
Key takeaway 24 · Languages

We assign four language tiers based on whether two solutions meet the requirements and how many steps they need.

We aim to get at least two solutions that meet all the requirements of the level (a main solution and a backup). In tier 1, at least two one-step solutions meet them.

In tier 2, at least two meet them only as two-step solutions (translation + post-edit), three-step solutions (translation + LQA + repair) or, for gender neutrality, as solutions with the gender-neutrality agent.

In tier 3, fewer than two solutions meet them after every step we added. Tier 4 holds the languages where the automatic metrics or Intento LQA cannot assess quality.

See details →  ·  Explore interactively →  ·  See more on Languages →

Errors, cost and time by language tier, for each quality levelTier 1: one-step solutionsTier 2: complex solutionsTier 3: no reliable high qualitymiddle half of the languages (25th to 75th percentile)Errors per 1,000 words, best solution0510152025Basic, Tier 1: one-step solutions: 11 (median of 41 languages, middle half 10 to 14)41Basic, Tier 2: complex solutions: 9 (median of 38 languages, middle half 8 to 10)38Basic, Tier 3: no reliable high quality: 19 (median of 3 languages, middle half 14 to 20)3BasicStandard, Tier 1: one-step solutions: 11 (median of 20 languages, middle half 10 to 12)20Standard, Tier 2: complex solutions: 10 (median of 58 languages, middle half 8 to 12)58Standard, Tier 3: no reliable high quality: 16 (median of 4 languages, middle half 13 to 19)4StandardAdvanced, Tier 1: one-step solutions: 12 (median of 6 languages, middle half 10 to 12)6Advanced, Tier 2: complex solutions: 9 (median of 47 languages, middle half 7 to 11)47Advanced, Tier 3: no reliable high quality: 12 (median of 29 languages, middle half 7 to 15)29AdvancedCost per million characters, cheapest$1$10$100$1,000$10,000Basic, Tier 1: one-step solutions: $5 (median of 41 languages, middle half $3 to $7)41Basic, Tier 2: complex solutions: $546 (median of 38 languages, middle half $23 to $577)38Basic, Tier 3: no reliable high quality: $636 (median of 3 languages, middle half $599 to $659)3BasicStandard, Tier 1: one-step solutions: $4 (median of 20 languages, middle half $4 to $5)20Standard, Tier 2: complex solutions: $30 (median of 58 languages, middle half $9 to $567)58Standard, Tier 3: no reliable high quality: $599 (median of 4 languages, middle half $562 to $647)4StandardAdvanced, Tier 1: one-step solutions: $7 (median of 6 languages, middle half $1 to $10)6Advanced, Tier 2: complex solutions: $47 (median of 47 languages, middle half $10 to $596)47Advanced, Tier 3: no reliable high quality: $968 (median of 29 languages, middle half $63 to $1,032)29AdvancedSeconds per 1,000 words, fastest0100200300400500Basic, Tier 1: one-step solutions: 55 (median of 41 languages, middle half 24 to 74)41Basic, Tier 2: complex solutions: 178 (median of 38 languages, middle half 118 to 257)38Basic, Tier 3: no reliable high quality: 243 (median of 3 languages, middle half 223 to 458)3BasicStandard, Tier 1: one-step solutions: 49 (median of 20 languages, middle half 26 to 64)20Standard, Tier 2: complex solutions: 137 (median of 58 languages, middle half 74 to 208)58Standard, Tier 3: no reliable high quality: 262 (median of 4 languages, middle half 233 to 379)4StandardAdvanced, Tier 1: one-step solutions: 35 (median of 6 languages, middle half 16 to 49)6Advanced, Tier 2: complex solutions: 191 (median of 47 languages, middle half 112 to 254)47Advanced, Tier 3: no reliable high quality: 224 (median of 29 languages, middle half 191 to 300)29AdvancedUnder each bar, the number of languages in the tier.Medians over the languages of the tier. In tier 3, the best solution of the language. The cost and time of three-step solutions include the LQA step, an estimate. Tier 4 (18languages) has no LQA result.
Errors, cost and time by language tier, for each quality level

In each cell: the languages of the tier, and over them the median total errors per 1,000 words (EPT) of the best solution across all requirements, the median seconds per 1,000 words of the fastest and the median cost per million characters of the cheapest, with the distance to the 25th and 75th percentiles in small type. Where no solution meets the selected requirements, the values of the best available solution.

Tier 1: one-step solutionsTier 2: complex solutionsTier 3: no reliable high qualityTier 4: cannot assessBasic requirements
41 languages16.5 EPT+7.2−3.155 sec+20−31$4.61/1M+$2.84−$1.15
38 languages11.0 EPT+2.1−2.9178 sec+79−59$546/1M+$32−$522
3 languages19.1 EPT+1.1−3.3243 sec+215−20$636/1M+$23−$37
18 languagescannot assess
Standard requirements
20 languages16.1 EPT+6.7−3.649 sec+15−24$4.50/1M+$0.98−$0.49
58 languages13.0 EPT+4.2−3.6137 sec+71−63$30/1M+$537−$21
4 languages16.8 EPT+2.9−2.8262 sec+117−29$599/1M+$48−$37
18 languagescannot assess
Advanced requirements
6 languages11.6 EPT+0.9−1.435 sec+14−19$7.10/1M+$3.26−$5.78
47 languages8.9 EPT+1.8−2.0191 sec+63−80$47/1M+$549−$37
29 languages12.2 EPT+2.6−4.3225 sec+76−33$968/1M+$64−$905
18 languagescannot assess
Key takeaway 25 · Languages

No low-resource language has two one-step solutions meeting Basic requirements, and 11 of 40 high-resource languages also need more steps.

They are Chinese Simplified and Traditional, Estonian, Finnish, French, Hebrew, Hungarian, Japanese, Korean, Latvian and Thai. Languages of one family often share a tier: all 12 Slavic languages are in tier 1 for Basic, and the Dravidian languages (Kannada, Malayalam, Tamil, Telugu) and the Uralic languages (Estonian, Finnish, Hungarian) are in tier 2 at every level. All 18 languages we cannot assess are low-resource or medium-resource languages.

See more on Languages →

Language tiers for the Basic requirements, by resource levelTier 1: one-step solutionsTier 2: complex solutionsTier 3: no reliable high qualityTier 4: cannot assess010203040High-resource languages (40)High-resource languages (40): Tier 1: one-step solutions 29 (72.5%)29High-resource languages (40): Tier 2: complex solutions 11 (27.5%)11High-resource languages (40): Tier 3: no reliable high quality 0 (0.0%)High-resource languages (40): Tier 4: cannot assess 0 (0.0%)Medium-resource languages (35)Medium-resource languages (35): Tier 1: one-step solutions 12 (34.3%)12Medium-resource languages (35): Tier 2: complex solutions 17 (48.6%)17Medium-resource languages (35): Tier 3: no reliable high quality 2 (5.7%)2Medium-resource languages (35): Tier 4: cannot assess 4 (11.4%)4Low-resource languages (25)Low-resource languages (25): Tier 1: one-step solutions 0 (0.0%)Low-resource languages (25): Tier 2: complex solutions 10 (40.0%)10Low-resource languages (25): Tier 3: no reliable high quality 1 (4.0%)Low-resource languages (25): Tier 4: cannot assess 14 (56.0%)14languages100 languages: 40 high-resource, 35 medium-resource and 25 low-resource languages, resource level asin the language features of 00-07.
Language tiers for the Basic requirements, by resource level
Key takeaway 26 · Languages

In 13 languages, full gender neutrality requires rewording because forms of address or verb agreement have no neutral alternative.

In Hindi, Urdu, Punjabi and Gujarati, the verb agrees with the gender of the person, also with the polite form of you. In Croatian, Serbian, Bosnian, Slovenian, Czech, Polish and Lithuanian, past-tense forms and participles carry gender after the polite you or the authorial we. In Arabic and Hebrew, all forms of address are masculine or feminine. We assess these languages against the same gender rule as every other language, and in 12 of them at most one of the two best solutions is fully gender-neutral.

See more on Languages →

Key takeaway 27 · Quality

We set pass-or-fail thresholds for nine requirements and report locale conventions as advisory.

For general translation quality, we weight every error by severity, 1 for a minor error, 5 for a major and 10 for a critical, and a translation meets the bar with no critical error, at most three major errors in the four texts and at most 20 weighted errors per 1,000 words. That leaves room for up to 60 minor errors in the ~3,000 words.

We set 100% for the other Standard requirements, because we check each one as met or not met, by code for the formatting requirements and sentence by sentence for tone of voice and terminology, and because a broken tag or an invalid JSON document can stop a translation from being imported or displayed correctly.

Terminology applies in the 16 languages with a glossary, and Unicode character integrity includes look-alike letters from another script. We report locale conventions as a share without a bar, because whether to localize punctuation is a matter of house style. For gender neutrality (Advanced), we also require 100% of the sentences that refer to a person.

See more on Quality →

Thresholds for the requirements at each quality level1. BasicGeneral translation qualityNo critical error, at most 3 majorerrors and at most 20 weightederrors per 1,000 words2. StandardEverything in Basic, plusTerminology100%, in the 16 languages with aglossaryTone of voice100%Tags100%Unicode character integrity100%, including look-alike lettersfrom another scriptLocale conventionsNo bar, reported as a shareHTML entities100%Valid JSON100%Spacing at tag boundaries100%3. AdvancedEverything in Standard, plusGender neutrality100% of the sentences that refer toa personWeighted errors: minor 1, major 5, critical 10, over the four texts of ~3,000 words.
Thresholds for the requirements at each quality level
Key takeaway 28 · Quality

Fewer than two solutions meet Basic requirements in 3 of 82 languages, Standard in 4 and Advanced in 29.

For Basic, Afrikaans has one such solution, Burmese and Sundanese none. For Standard, Afrikaans and Frisian have one, Burmese and Sundanese none. For Advanced, 11 languages have none: Albanian, Arabic, Burmese, Czech, French, Hebrew, Polish, Sinhala, Sundanese, Urdu and Yiddish. The other 18 have one.

See details →  ·  See more on Quality →

Languages by the number of solutions meeting each quality levelTwo or more solutionsOne solutionNo solution020406080100BasicBasic: Two or more solutions 79 (96.3%)Basic: One solution 1 (1.2%)Basic: No solution 2 (2.4%)StandardStandard: Two or more solutions 78 (95.1%)Standard: One solution 2 (2.4%)Standard: No solution 2 (2.4%)AdvancedAdvanced: Two or more solutions 53 (64.6%)Advanced: One solution 18 (22.0%)Advanced: No solution 11 (13.4%)languages82 languages. Basic and Standard: solutions up to three steps. Advanced: the two best solutions of each language, with thegender-neutrality agent.
Languages by the number of solutions meeting each quality level
Key takeaway 29 · Quality

Accuracy errors account for 60% of severity-weighted translation errors, including 38% from mistranslation.

Fluency errors account for 36% of severity-weighted translation errors.

Accuracy errors are mistranslation, omission, addition, under- and over-translation and untranslated text, and fluency errors are grammar, spelling and punctuation. Grammar is the second largest type, at 20%, and the other 4% are broken characters, stray markup and corrupted output.

See more on Quality →

General translation quality errors by type0%10%20%30%AccuracyFluencyBroken characters or markupMistranslation: 38.1% of the weighted errorsMistranslation38.1%Grammar: 19.8% of the weighted errorsGrammar19.8%Punctuation: 9.4% of the weighted errorsPunctuation9.4%Addition: 7.1% of the weighted errorsAddition7.1%Omission: 6.9% of the weighted errorsOmission6.9%Spelling: 6.6% of the weighted errorsSpelling6.6%Untranslated text: 2.8% of the weighted errorsUntranslated text2.8%Under-translation: 2.7% of the weighted errorsUnder-translation2.7%Over-translation: 2% of the weighted errorsOver-translation2%Character encoding: 1.7% of the weighted errorsCharacter encoding1.7%Stray markup: 1.6% of the weighted errorsStray markup1.6%Corrupted output: 1.1% of the weighted errorsCorrupted output1.1%Culture-specific reference: 0.2% of the weighted errorsCulture-specific reference0.2%% of the weighted errorsShare of the weighted errors in the translations of the one-step solutions with an Intento LQA result in 82 languages.Weighted errors: minor 1, major 5, critical 10.
General translation quality errors by type
Key takeaway 30 · Quality

LLM translations meet general translation quality requirements in 21% of cases, while MT engine translations almost never do.

Among the evaluated LLM solutions that initially fall short, 24% meet this requirement after two steps and 60% after three. General translation quality covers translation errors of every type, such as mistranslations, omissions and grammar errors, which we classify by type and severity following MQM, the industry standard for classifying translation errors.

A solution meets it with no critical error, at most three major errors and at most 20 weighted errors per 1,000 words in the four texts (minor 1, major 5, critical 10). Translations from OpenAI GPT-5.6 Sol meet it in 20 of the 32 languages where we shortlisted them and from Anthropic Claude Fable 5 in 33, and among MT engines only translations from DeepL Next Gen do, in German and Romanian.

In 28 languages, among them French, Japanese and Korean, every one-step solution misses it. Where an LLM solution misses it, the median goes from 36 weighted errors per 1,000 words to 28 in the two-step solution and to 16 in the three-step solution, below the bar. We evaluated these repairs on LLM solutions. MT engine solutions fell short on too many requirements to enter this repair comparison.

See details →  ·  See more on Quality →

Weighted errors per 1,000 words of one-step solutions, median by modelMT enginesLLM-based MT enginesLLMsOpen-weight LLMs020406080100OpenAI GPT-5.6 Sol: 18.618.6OpenAI GPT-5.6 SolAnthropic Claude Fable 5: 19.419.4Anthropic Claude Fable 5OpenAI GPT-5.6 Terra: 20.620.6OpenAI GPT-5.6 TerraOpenAI GPT-5.5: 22.122.1OpenAI GPT-5.5Anthropic Claude Opus 4.8: 24.424.4Anthropic Claude Opus 4.8Google Gemini 3.1 Pro Preview: 2525Google Gemini 3.1 Pro PreviewOpenAI GPT-5.6 Luna: 28.928.9OpenAI GPT-5.6 LunaCohere Command A Translate: 33.233.2Cohere Command A TranslateAnthropic Claude Sonnet 5: 35.335.3Anthropic Claude Sonnet 5Anthropic Claude Sonnet 4.6: 35.635.6Anthropic Claude Sonnet 4.6Google Gemini 3.5 Flash: 35.935.9Google Gemini 3.5 FlashMoonshotAI Kimi K2.5: 36.636.6MoonshotAI Kimi K2.5Translated Lara 2: 38.238.2Translated Lara 2OpenAI GPT-4.1: 38.238.2OpenAI GPT-4.1xAI Grok 4.20 (reasoning): 41.241.2xAI Grok 4.20 (reasoning)DeepL Next Gen: 41.941.9DeepL Next Genz.ai GLM-5: 43.243.2z.ai GLM-5Mistral Large 3: 45.345.3Mistral Large 3Google Gemma 4: 47.847.8Google Gemma 4DeepSeek V3.2: 52.252.2DeepSeek V3.2OpenAI gpt-oss-120b: 58.858.8OpenAI gpt-oss-120bWidn AI Tower Sugarloaf: 60.260.2Widn AI Tower SugarloafxAI Grok 4.20 (non-reasoning): 65.665.6xAI Grok 4.20 (non-reasoning)Microsoft Translator: 67.967.9Microsoft TranslatorGoogle Translation LLM: 81.981.9Google Translation LLMModels on the shortlist of eight or more of the 82 languages, one-step solutions. Minor 1, major 5, critical 10. Generaltranslation quality: at most 20, no critical error and at most three major errors in the four texts.Company, product and model names and logos are shown for informational purposes only. All trademarks andlogos belong to their respective owners. Their use does not imply affiliation or endorsement.
Weighted errors per 1,000 words of one-step solutions, median by model
Key takeaway 31 · Quality

LLM translations meet glossary requirements more often than MT engine translations, and additional steps can correct terminology errors.

Terminology requires the approved translation of every glossary term, in any of its correct forms.

In the 16 languages with a glossary, translations from LLMs have terminology errors in 43% of the cases, translations from LLM-based MT engines in 68% and translations from other MT engines in every case. Translations from Google Gemini 3.1 Pro Preview use every term in all 15 glossary languages where we shortlisted them, and translations from Microsoft Translator have a median of 13.9 terminology errors per 1,000 words.

Translations into Chinese Simplified and Ukrainian have the most, a median of 1.1 and 0.8 terminology errors per 1,000 words. Two-step and three-step solutions fix nearly every case, though the repair in three-step solutions can also paraphrase a term that was right. We evaluated these repairs on LLM solutions. MT engine solutions fell short on too many requirements to enter this repair comparison.

See details →  ·  See more on Quality →

Terminology errors per 1,000 words, one-step solutions, median by modelMT enginesLLM-based MT enginesLLMs051015Google Gemini 3.1 Pro Preview: 00Google Gemini 3.1 Pro PreviewAnthropic Claude Fable 5: 00Anthropic Claude Fable 5OpenAI GPT-4.1: 00OpenAI GPT-4.1Anthropic Claude Sonnet 4.6: 00Anthropic Claude Sonnet 4.6Anthropic Claude Opus 4.8: 00Anthropic Claude Opus 4.8Google Gemini 3.5 Flash: 00Google Gemini 3.5 FlashDeepL Next Gen: 0.50.5DeepL Next GenAnthropic Claude Sonnet 5: 0.80.8Anthropic Claude Sonnet 5Cohere Command A Translate: 11Cohere Command A TranslateMicrosoft Translator: 13.913.9Microsoft TranslatorModels on the shortlist of eight or more of the 16 languages with a glossary, one-step solutions. 121 glossary terms in thefour texts, declined forms accepted.Company, product and model names and logos are shown for informational purposes only. All trademarks andlogos belong to their respective owners. Their use does not imply affiliation or endorsement.
Terminology errors per 1,000 words, one-step solutions, median by model
Key takeaway 32 · Quality

Register errors occur in 52% of MT engine translations and 9% of LLM translations.

Among the evaluated LLM solutions with register errors, 68% meet the requirement after two steps and 73% after three. Tone of voice requires the formal register in every sentence in the second person or the imperative, such as Sie in German and 敬語 in Japanese.

Translations from LLM-based MT engines have register errors in 28% of the cases. Translations from Microsoft Translator have the most, a mean of 3.8 register errors per 1,000 words, and translations from Anthropic Claude Fable 5, Anthropic Claude Sonnet 5 and z.ai GLM-5 a mean below 0.05.

Translations into Cebuano, Finnish and Sundanese have the most, a mean of 7.8, 5.3 and 3.9 register errors per 1,000 words, and translations into 51 of the 82 languages have no register errors. We evaluated these repairs on LLM solutions. MT engine solutions fell short on too many requirements to enter this repair comparison.

See details →  ·  See more on Quality →

Register errors per 1,000 words, one-step solutions, mean by modelMT enginesLLM-based MT enginesLLMsOpen-weight LLMs01234Anthropic Claude Fable 5: 00Anthropic Claude Fable 5Anthropic Claude Sonnet 5: 00Anthropic Claude Sonnet 5z.ai GLM-5: 00z.ai GLM-5OpenAI GPT-5.6 Sol: 00OpenAI GPT-5.6 SolMistral Large 3: 00Mistral Large 3xAI Grok 4.20 (non-reasoning): 00xAI Grok 4.20 (non-reasoning)Google Gemini 3.5 Flash: 0.10.1Google Gemini 3.5 FlashOpenAI GPT-5.5: 0.10.1OpenAI GPT-5.5Anthropic Claude Opus 4.8: 0.10.1Anthropic Claude Opus 4.8OpenAI GPT-4.1: 0.10.1OpenAI GPT-4.1Anthropic Claude Sonnet 4.6: 0.10.1Anthropic Claude Sonnet 4.6OpenAI GPT-5.6 Luna: 0.10.1OpenAI GPT-5.6 LunaOpenAI GPT-5.6 Terra: 0.10.1OpenAI GPT-5.6 TerraDeepSeek V3.2: 0.10.1DeepSeek V3.2Google Gemini 3.1 Pro Preview: 0.20.2Google Gemini 3.1 Pro PreviewTranslated Lara 2: 0.20.2Translated Lara 2xAI Grok 4.20 (reasoning): 0.20.2xAI Grok 4.20 (reasoning)MoonshotAI Kimi K2.5: 0.20.2MoonshotAI Kimi K2.5Google Gemma 4: 0.40.4Google Gemma 4Cohere Command A Translate: 0.70.7Cohere Command A TranslateOpenAI gpt-oss-120b: 1.11.1OpenAI gpt-oss-120bGoogle Translation LLM: 1.31.3Google Translation LLMDeepL Next Gen: 1.41.4DeepL Next GenWidn AI Tower Sugarloaf: 1.91.9Widn AI Tower SugarloafMicrosoft Translator: 3.83.8Microsoft TranslatorModels on the shortlist of eight or more of the 82 languages, one-step solutions. A register error is a sentence in thesecond person or the imperative that is not in the formal register (47 such sentences in the four texts). The mean, sincethe median is 0 for nearly every model.Company, product and model names and logos are shown for informational purposes only. All trademarks andlogos belong to their respective owners. Their use does not imply affiliation or endorsement.
Register errors per 1,000 words, one-step solutions, mean by model
Key takeaway 33 · Quality

Tag errors occur in 98% of MT engine translations and 41% of LLM translations.

Where tags are the only problem, code repair restores them in 90% of the evaluated cases. Tag handling requires every HTML tag of the source, with its attributes, in the translation, since a broken tag can stop a translation from being imported or displayed correctly.

Translations from LLM-based MT engines lose tags in 80% of the cases. Translations from Google Gemini 3.1 Pro Preview keep every tag in 72 of the 75 languages where we shortlisted them, translations from Anthropic Claude Opus 4.8 in 11 of 45, and translations from DeepL Next Gen lose tags in every language where we shortlisted them.

Translations into European Spanish, Japanese, Chinese Simplified and Marathi have the most tag errors, a median of 2.0 to 2.3 per 1,000 words. Where general translation quality also falls short, two-step solutions (translation + post-edit) restore the tags in 59% of the cases and three-step solutions (translation + LQA + repair) in 56%. We evaluated these repairs on LLM solutions. MT engine solutions fell short on too many requirements to enter this repair comparison.

See details →  ·  See more on Quality →

Tag errors per 1,000 words, one-step solutions, median by modelMT enginesLLM-based MT enginesLLMsOpen-weight LLMs0204060Google Gemini 3.1 Pro Preview: 00Google Gemini 3.1 Pro PreviewGoogle Gemini 3.5 Flash: 00Google Gemini 3.5 FlashAnthropic Claude Fable 5: 00Anthropic Claude Fable 5OpenAI GPT-5.5: 00OpenAI GPT-5.5Translated Lara 2: 00Translated Lara 2OpenAI GPT-4.1: 00OpenAI GPT-4.1z.ai GLM-5: 00z.ai GLM-5OpenAI GPT-5.6 Sol: 00OpenAI GPT-5.6 SolOpenAI GPT-5.6 Terra: 00OpenAI GPT-5.6 TerraDeepSeek V3.2: 00DeepSeek V3.2OpenAI GPT-5.6 Luna: 0.70.7OpenAI GPT-5.6 LunaxAI Grok 4.20 (reasoning): 0.70.7xAI Grok 4.20 (reasoning)MoonshotAI Kimi K2.5: 0.70.7MoonshotAI Kimi K2.5OpenAI gpt-oss-120b: 11OpenAI gpt-oss-120bCohere Command A Translate: 1.21.2Cohere Command A TranslateAnthropic Claude Opus 4.8: 1.31.3Anthropic Claude Opus 4.8Google Gemma 4: 1.31.3Google Gemma 4Anthropic Claude Sonnet 5: 1.31.3Anthropic Claude Sonnet 5Anthropic Claude Sonnet 4.6: 1.31.3Anthropic Claude Sonnet 4.6Mistral Large 3: 1.31.3Mistral Large 3xAI Grok 4.20 (non-reasoning): 1.31.3xAI Grok 4.20 (non-reasoning)Microsoft Translator: 5.95.9Microsoft TranslatorGoogle Translation LLM: 8.28.2Google Translation LLMDeepL Next Gen: 9.29.2DeepL Next GenWidn AI Tower Sugarloaf: 59.859.8Widn AI Tower SugarloafModels on the shortlist of eight or more of the 82 languages, one-step solutions. A tag error is a tag of the source lostor changed, checked with its attributes (508 tags in the four texts).Company, product and model names and logos are shown for informational purposes only. All trademarks andlogos belong to their respective owners. Their use does not imply affiliation or endorsement.
Tag errors per 1,000 words, one-step solutions, median by model
Key takeaway 34 · Quality

Unicode errors occur in 3% to 6% of LLM and MT engine translations, compared with 16% for open-weight LLMs.

Unicode character integrity requires every character of the translation intact, with no corrupted characters and no letter from another script that only looks like the right one.

Look-alike letters from another script occur most often in Kyrgyz, Tatar and Macedonian, and in Slovak, translations from Google Gemini 3.5 Flash, z.ai GLM-5, DeepL Next Gen and Google Translation LLM carry corrupted characters. Translations from Google Gemma 4 keep Unicode characters intact in 30 of 37 languages, and translations from 9 of the 25 models on eight or more shortlists in every language.

Three-step solutions (translation + LQA + repair) fix every Unicode error of the one-step solutions we built them on. We evaluated these repairs on LLM solutions. MT engine solutions fell short on too many requirements to enter this repair comparison.

See details →  ·  See more on Quality →

Languages with and without Unicode errors, by modelLanguages with Unicode characters intactOther languages where we shortlisted the model020406080Google Gemini 3.1 Pro Preview: Languages with Unicode characters intact 75 of 7575Google Gemini 3.1 Pro PreviewDeepL Next Gen: Languages with Unicode characters intact 66 of 69DeepL Next Gen: Other languages where we shortlisted the model 3 of 6969DeepL Next GenGoogle Gemini 3.5 Flash: Languages with Unicode characters intact 62 of 63Google Gemini 3.5 Flash: Other languages where we shortlisted the model 1 of 6363Google Gemini 3.5 FlashAnthropic Claude Fable 5: Languages with Unicode characters intact 60 of 6060Anthropic Claude Fable 5OpenAI GPT-5.5: Languages with Unicode characters intact 57 of 59OpenAI GPT-5.5: Other languages where we shortlisted the model 2 of 5959OpenAI GPT-5.5Anthropic Claude Opus 4.8: Languages with Unicode characters intact 44 of 45Anthropic Claude Opus 4.8: Other languages where we shortlisted the model 1 of 4545Anthropic Claude Opus 4.8OpenAI GPT-4.1: Languages with Unicode characters intact 43 of 4343OpenAI GPT-4.1Translated Lara 2: Languages with Unicode characters intact 43 of 44Translated Lara 2: Other languages where we shortlisted the model 1 of 4444Translated Lara 2Anthropic Claude Sonnet 5: Languages with Unicode characters intact 33 of 35Anthropic Claude Sonnet 5: Other languages where we shortlisted the model 2 of 3535Anthropic Claude Sonnet 5OpenAI GPT-5.6 Sol: Languages with Unicode characters intact 32 of 3232OpenAI GPT-5.6 SolMicrosoft Translator: Languages with Unicode characters intact 31 of 32Microsoft Translator: Other languages where we shortlisted the model 1 of 3232Microsoft Translatorz.ai GLM-5: Languages with Unicode characters intact 31 of 33z.ai GLM-5: Other languages where we shortlisted the model 2 of 3333z.ai GLM-5Google Gemma 4: Languages with Unicode characters intact 30 of 37Google Gemma 4: Other languages where we shortlisted the model 7 of 3737Google Gemma 4Anthropic Claude Sonnet 4.6: Languages with Unicode characters intact 29 of 31Anthropic Claude Sonnet 4.6: Other languages where we shortlisted the model 2 of 3131Anthropic Claude Sonnet 4.6Mistral Large 3: Languages with Unicode characters intact 27 of 28Mistral Large 3: Other languages where we shortlisted the model 1 of 2828Mistral Large 3Google Translation LLM: Languages with Unicode characters intact 27 of 32Google Translation LLM: Other languages where we shortlisted the model 5 of 3232Google Translation LLMOpenAI GPT-5.6 Luna: Languages with Unicode characters intact 26 of 2626OpenAI GPT-5.6 LunaCohere Command A Translate: Languages with Unicode characters intact 21 of 22Cohere Command A Translate: Other languages where we shortlisted the model 1 of 2222Cohere Command A TranslateOpenAI GPT-5.6 Terra: Languages with Unicode characters intact 18 of 20OpenAI GPT-5.6 Terra: Other languages where we shortlisted the model 2 of 2020OpenAI GPT-5.6 TerraxAI Grok 4.20 (reasoning): Languages with Unicode characters intact 15 of 1515xAI Grok 4.20 (reasoning)DeepSeek V3.2: Languages with Unicode characters intact 14 of 1414DeepSeek V3.2OpenAI gpt-oss-120b: Languages with Unicode characters intact 12 of 14OpenAI gpt-oss-120b: Other languages where we shortlisted the model 2 of 1414OpenAI gpt-oss-120bMoonshotAI Kimi K2.5: Languages with Unicode characters intact 11 of 1111MoonshotAI Kimi K2.5Widn AI Tower Sugarloaf: Languages with Unicode characters intact 10 of 1010Widn AI Tower SugarloafxAI Grok 4.20 (non-reasoning): Languages with Unicode characters intact 6 of 8xAI Grok 4.20 (non-reasoning): Other languages where we shortlisted the model 2 of 88xAI Grok 4.20 (non-reasoning)Models on the shortlist of eight or more of the 82 languages, one-step solutions, in order of the languages with Unicodecharacters intact, most first. Intact: no corrupted character in the four texts and no look-alike letter from another script.Company, product and model names and logos are shown for informational purposes only. All trademarks and logosbelong to their respective owners. Their use does not imply affiliation or endorsement.
Languages with and without Unicode errors, by model
Key takeaway 35 · Quality

The translated JSON document is invalid in 92% of MT engine cases and 3% of LLM cases.

All evaluated two-step and three-step repairs of invalid JSON produced valid documents. Valid JSON requires the translated JSON document to still parse, since an invalid JSON document can stop a translation from being imported.

Translations from LLM-based MT engines have an invalid JSON document in 74% of the cases, and translations from Translated Lara 2, Microsoft Translator, Google Translation LLM, Widn AI Tower Sugarloaf and Amazon Translate in every language where we shortlisted them. Among LLMs, translations from Anthropic Claude Sonnet 4.6 have one in 10 of 31 languages, and translations from open-weight LLMs always return a valid one.

In Marathi and Georgian, most of the invalid documents come from the LLM-based MT engines on the shortlist. We evaluated these repairs on LLM solutions. MT engine solutions fell short on too many requirements to enter this repair comparison.

See details →  ·  See more on Quality →

Languages with valid and invalid translated JSON, by modelLanguages with a valid JSON documentOther languages where we shortlisted the model020406080Google Gemini 3.1 Pro Preview: Languages with a valid JSON document 75 of 7575Google Gemini 3.1 Pro PreviewGoogle Gemini 3.5 Flash: Languages with a valid JSON document 62 of 63Google Gemini 3.5 Flash: Other languages where we shortlisted the model 1 of 6363Google Gemini 3.5 FlashOpenAI GPT-5.5: Languages with a valid JSON document 59 of 5959OpenAI GPT-5.5Anthropic Claude Fable 5: Languages with a valid JSON document 59 of 60Anthropic Claude Fable 5: Other languages where we shortlisted the model 1 of 6060Anthropic Claude Fable 5Anthropic Claude Opus 4.8: Languages with a valid JSON document 45 of 4545Anthropic Claude Opus 4.8OpenAI GPT-4.1: Languages with a valid JSON document 42 of 43OpenAI GPT-4.1: Other languages where we shortlisted the model 1 of 4343OpenAI GPT-4.1DeepL Next Gen: Languages with a valid JSON document 40 of 69DeepL Next Gen: Other languages where we shortlisted the model 29 of 6969DeepL Next GenGoogle Gemma 4: Languages with a valid JSON document 37 of 3737Google Gemma 4Anthropic Claude Sonnet 5: Languages with a valid JSON document 35 of 3535Anthropic Claude Sonnet 5OpenAI GPT-5.6 Sol: Languages with a valid JSON document 32 of 3232OpenAI GPT-5.6 Solz.ai GLM-5: Languages with a valid JSON document 32 of 33z.ai GLM-5: Other languages where we shortlisted the model 1 of 3333z.ai GLM-5Mistral Large 3: Languages with a valid JSON document 27 of 28Mistral Large 3: Other languages where we shortlisted the model 1 of 2828Mistral Large 3OpenAI GPT-5.6 Luna: Languages with a valid JSON document 26 of 2626OpenAI GPT-5.6 LunaCohere Command A Translate: Languages with a valid JSON document 21 of 22Cohere Command A Translate: Other languages where we shortlisted the model 1 of 2222Cohere Command A TranslateAnthropic Claude Sonnet 4.6: Languages with a valid JSON document 21 of 31Anthropic Claude Sonnet 4.6: Other languages where we shortlisted the model 10 of 3131Anthropic Claude Sonnet 4.6OpenAI GPT-5.6 Terra: Languages with a valid JSON document 20 of 2020OpenAI GPT-5.6 TerraxAI Grok 4.20 (reasoning): Languages with a valid JSON document 15 of 1515xAI Grok 4.20 (reasoning)OpenAI gpt-oss-120b: Languages with a valid JSON document 14 of 1414OpenAI gpt-oss-120bDeepSeek V3.2: Languages with a valid JSON document 13 of 14DeepSeek V3.2: Other languages where we shortlisted the model 1 of 1414DeepSeek V3.2MoonshotAI Kimi K2.5: Languages with a valid JSON document 10 of 11MoonshotAI Kimi K2.5: Other languages where we shortlisted the model 1 of 1111MoonshotAI Kimi K2.5xAI Grok 4.20 (non-reasoning): Languages with a valid JSON document 7 of 8xAI Grok 4.20 (non-reasoning): Other languages where we shortlisted the model 1 of 88xAI Grok 4.20 (non-reasoning)Widn AI Tower Sugarloaf: Other languages where we shortlisted the model 10 of 1010Widn AI Tower SugarloafGoogle Translation LLM: Other languages where we shortlisted the model 32 of 3232Google Translation LLMMicrosoft Translator: Other languages where we shortlisted the model 32 of 3232Microsoft TranslatorTranslated Lara 2: Other languages where we shortlisted the model 44 of 4444Translated Lara 2Models on the shortlist of eight or more of the 82 languages, one-step solutions, in order of the languages with a valid JSONdocument, most first. Valid: the translation of the JSON text still parses as JSON.Company, product and model names and logos are shown for informational purposes only. All trademarks and logosbelong to their respective owners. Their use does not imply affiliation or endorsement.
Languages with valid and invalid translated JSON, by model
Key takeaway 36 · Quality

Gender-neutrality errors have a median rate of 5.1 per 1,000 words in gendered and natural-gender languages, compared with 0 in genderless languages.

Gender neutrality requires a gender-neutral translation wherever the English source leaves the gender of a person open, as in "your manager" or "the expert", which Spanish, German and Arabic would mark.

Two or more one-step solutions are fully gender-neutral in 34 of the 82 languages, 23 of them genderless languages, which have no gendered nouns or pronouns. In 46 languages, among them the Romance and Slavic languages, German, Greek, Arabic and Hebrew, translations from every shortlisted model have gender-neutrality errors. By type, translations from LLMs have them in 67% of the cases and translations from MT engines in 86%.

Where the two best solutions of a language are not fully gender-neutral, we add the gender-neutrality agent, which makes 47% of the one-step versions fully gender-neutral, 54% of the two-step versions (translation + post-edit) and 68% of the three-step versions (translation + LQA + repair).

See details →  ·  See more on Quality →

Gender-neutrality errors per 1,000 words, one-step solutions, median by modelMT enginesLLM-based MT enginesLLMsOpen-weight LLMs02.557.51012.5Google Translation LLM: 0.30.3Google Translation LLMOpenAI GPT-5.5: 0.40.4OpenAI GPT-5.5Anthropic Claude Opus 4.8: 0.40.4Anthropic Claude Opus 4.8MoonshotAI Kimi K2.5: 0.40.4MoonshotAI Kimi K2.5xAI Grok 4.20 (non-reasoning): 0.40.4xAI Grok 4.20 (non-reasoning)OpenAI GPT-5.6 Luna: 0.50.5OpenAI GPT-5.6 LunaDeepSeek V3.2: 0.70.7DeepSeek V3.2Google Gemini 3.5 Flash: 11Google Gemini 3.5 Flashz.ai GLM-5: 1.11.1z.ai GLM-5Anthropic Claude Fable 5: 1.21.2Anthropic Claude Fable 5Google Gemini 3.1 Pro Preview: 1.31.3Google Gemini 3.1 Pro PreviewxAI Grok 4.20 (reasoning): 1.41.4xAI Grok 4.20 (reasoning)OpenAI gpt-oss-120b: 1.51.5OpenAI gpt-oss-120bGoogle Gemma 4: 22Google Gemma 4DeepL Next Gen: 2.32.3DeepL Next GenMistral Large 3: 2.62.6Mistral Large 3Translated Lara 2: 2.72.7Translated Lara 2OpenAI GPT-5.6 Sol: 2.82.8OpenAI GPT-5.6 SolOpenAI GPT-4.1: 33OpenAI GPT-4.1Anthropic Claude Sonnet 5: 3.13.1Anthropic Claude Sonnet 5OpenAI GPT-5.6 Terra: 3.23.2OpenAI GPT-5.6 TerraAnthropic Claude Sonnet 4.6: 3.83.8Anthropic Claude Sonnet 4.6Microsoft Translator: 4.64.6Microsoft TranslatorWidn AI Tower Sugarloaf: 8.78.7Widn AI Tower SugarloafCohere Command A Translate: 11.811.8Cohere Command A TranslateModels on the shortlist of eight or more of the 82 languages, one-step solutions. A gender-neutrality error is a sentencethat refers to a person and is not gender-neutral (115 such sentences in the four texts). The two best solutions in eachlanguage use the rebuilt gender judge; all other solutions use the earlier judge.Company, product and model names and logos are shown for informational purposes only. All trademarks andlogos belong to their respective owners. Their use does not imply affiliation or endorsement.
Gender-neutrality errors per 1,000 words, one-step solutions, median by model
Key takeaway 37 · Costs

We calculate translation costs using August 2026 list prices and the measured tokens and characters.

Vendors sell translation per million tokens, per million characters, per word, per page, per request or by monthly subscription. We convert every price to US dollars per million characters of the English source, markup included, counting tokens with the tokenizer of each model. The four texts have 25,249 characters and 2,992 words, ~8.4 characters per word.

We take list prices from the vendor site, also for the models we used for free or on trial keys. For five models that are no longer on the price page, we use the last price the vendor listed, and for two open-weight models without a vendor API, the lowest price of an inference provider.

For 10 of the 55 models in our price check, there is no public price per token or per character. For seven of them, there is a quote or a sales contact instead of a price, a price shown only after sign-in, or no price list at all, and for three, the price is per word, per page or per request. Their solutions have no cost in our results.

For 11 models with reasoning, the cost leaves out the reasoning tokens, so it is a lower bound. For a two-step or three-step solution, we add up the cost of every step, the Intento LQA step included.

See details →  ·  See more on Costs →

Key takeaway 38 · Costs

Within the same language, MT engine solutions cost ~5 times as much as LLM solutions at the median and have more errors.

At the median, a solution on an MT engine costs $24.99 per million characters and on an LLM $4.59, and MT engines cost more than LLMs in all 96 languages. Prices run from $0.11 on the open-weight Google Gemma 4, at the lowest price of an inference provider, to $100 on ALL-Lab, an MT engine.

Translations from MT engines have a median 58 errors per 1,000 words weighted by severity, against 31 from LLMs, and more errors in 79 of the 84 languages with an Intento LQA result for both. 0% of the MT engine solutions meet every Standard requirement, against 12% of the LLM solutions.

Within each type, the price follows the quality. In 43 of the 46 languages with three or more MT engine solutions, the more expensive MT engines have fewer errors, as DeepL Next Gen and Translated Lara 2 against Microsoft Translator, and in 79 of 82 languages, the more expensive LLMs.

Per word, the median solution costs ~$0.00006, while in the 2025 edition we priced human translation at $0.16 to $0.24 per word, ~2,700 to 4,100 times as much.

See details →  ·  See more on Costs →

Cost and errors of the one-step solutions, by modelLLMsOpen-weight LLMsMT enginesLLM-based MT engines$0.1$1$10020406080100120US dollars per million characters, median over the languages (log scale)Weighted errors per 1,000 words, medianGoogle Gemma 4: $0.14 per million characters, 47.8 weighted errors per 1,000 words, 37 languagesOpenAI GPT-5.6 Luna: $0.45 per million characters, 28.9 weighted errors per 1,000 words, 26 languagesMistral Large 3: $0.82 per million characters, 45.3 weighted errors per 1,000 words, 28 languagesMistral Medium 3.1: $1.02 per million characters, 57.0 weighted errors per 1,000 words, 4 languagesDeepSeek V3.2: $1.21 per million characters, 52.2 weighted errors per 1,000 words, 14 languagesxAI Grok 4.20 (reasoning): $1.23 per million characters, 43.5 weighted errors per 1,000 words, 16 languagesMoonshotAI Kimi K2.5: $1.26 per million characters, 44.7 weighted errors per 1,000 words, 12 languagesxAI Grok 4.20 (non-reasoning): $1.28 per million characters, 67.6 weighted errors per 1,000 words, 9 languagesz.ai GLM-5: $2.4 per million characters, 43.2 weighted errors per 1,000 words, 33 languagesOpenAI GPT-4.1: $3.06 per million characters, 40.2 weighted errors per 1,000 words, 45 languagesGoogle Gemini 3.5 Flash: $3.57 per million characters, 35.9 weighted errors per 1,000 words, 64 languagesOpenAI GPT-5.6 Terra: $4.56 per million characters, 20.6 weighted errors per 1,000 words, 20 languagesGoogle Gemini 3.1 Pro Preview: $4.76 per million characters, 25.2 weighted errors per 1,000 words, 76 languagesAnthropic Claude Sonnet 5: $5.98 per million characters, 35.3 weighted errors per 1,000 words, 35 languagesAnthropic Claude Sonnet 4.6: $7.37 per million characters, 35.6 weighted errors per 1,000 words, 31 languagesOpenAI GPT-5.6 Sol: $7.9 per million characters, 18.8 weighted errors per 1,000 words, 33 languagesMicrosoft Translator: $10.0 per million characters, 67.9 weighted errors per 1,000 words, 32 languagesWidn AI Tower Sugarloaf: $10.0 per million characters, 60.2 weighted errors per 1,000 words, 10 languagesOpenAI GPT-5.5: $11.61 per million characters, 22.8 weighted errors per 1,000 words, 62 languagesAmazon: $15.0 per million characters, 76.1 weighted errors per 1,000 words, 7 languagesAnthropic Claude Opus 4.8: $15.83 per million characters, 25.4 weighted errors per 1,000 words, 46 languagesGoogle Cloud Advanced Translation API: $20.0 per million characters, 121.3 weighted errors per 1,000 words, 4 languagesGoogle Translation LLM: $21.32 per million characters, 81.9 weighted errors per 1,000 words, 32 languagesTranslated Lara 2: $24.99 per million characters, 38.2 weighted errors per 1,000 words, 44 languagesDeepL Next Gen: $27.5 per million characters, 42.8 weighted errors per 1,000 words, 71 languagesAnthropic Claude Fable 5: $32.28 per million characters, 19.4 weighted errors per 1,000 words, 60 languagesGoogle Cloud Advanced Translation APIGoogle Translation LLMAmazonMicrosoft TranslatorxAI Grok 4.20 (non-reasoning)Widn AI Tower SugarloafMistral Medium 3.1DeepSeek V3.2Google Gemma 4Mistral Large 3MoonshotAI Kimi K2.5xAI Grok 4.20 (reasoning)z.ai GLM-5DeepL Next GenOpenAI GPT-4.1Translated Lara 2Google Gemini 3.5 FlashAnthropic Claude Sonnet 4.6Anthropic Claude Sonnet 5OpenAI GPT-5.6 LunaAnthropic Claude Opus 4.8Google Gemini 3.1 Pro PreviewOpenAI GPT-5.5OpenAI GPT-5.6 TerraAnthropic Claude Fable 5OpenAI GPT-5.6 SolMedian over the languages where the model is on the shortlist with a conclusive Intento LQA result, models with three or more such languages. Errors weightedby severity (minor 1, major 5, critical 10). List prices of August 2026.Company, product and model names and logos are shown for informational purposes only. All trademarks and logos belong to theirrespective owners. Their use does not imply affiliation or endorsement.
Cost and errors of the one-step solutions, by model
Key takeaway 39 · Costs

A three-step solution costs a median of ~$590 per million source characters, mostly for Intento LQA.

The three-step median is ~30 times the two-step median and ~110 times the adjusted-LLM median. A one-step solution costs a median $27.50 per million characters on an off-the-shelf MT engine, LLM-based MT engines included, and $5.20 on an adjusted LLM. Tone of voice and a glossary in the request add ~2.5% to the cost of an LLM translation and nothing to an MT engine with a price per source character. A two-step solution (translation + post-edit) costs ~2.5 times as much as the one-step solution it starts from.

In a three-step solution (translation + LQA + repair), ~96% of the cost is the Intento LQA step, where we send every sentence to three LLM judges and their findings to a fourth LLM that combines them. A gender-neutral version costs a few percent more to translate than the solution it is a version of, and ~$1,000 as a three-step solution, gender checks included.

Each range covers several models and up to 96 languages. With billing on tokens, the same text costs more in languages with a longer translation and more tokens per character, so the ranges depend on the languages as well as on the models.

We use unified billing in Intento Language Hub to estimate costs across these model-and-harness workflows, including the Intento LQA step.

See details →  ·  See more on Costs →

Cost of the solutions per million characters, by level of solution complexityLowest to highestMiddle half of the solutionsMedian$0.1$1$10$100$1,000US dollars per million characters (log scale)Off-the-shelf NMTOff-the-shelf NMT: median $27.5, from $3.94 to $100.0 per million characters94 languagesAdjusted NMTAdjusted NMT: median $21.7, from $18.66 to $27.5 per million characters48 languagesOff-the-shelf LLMOff-the-shelf LLM: median $3.34, from $1.05 to $10.0 per million characters5 languagesAdjusted LLMAdjusted LLM: median $5.23, from $0.11 to $50.7 per million characters96 languagesTwo-step solutionTwo-step solution: median $18.28, from $0.28 to $107.12 per million characters64 languagesThree-step solutionThree-step solution: median $593.44, from $489.93 to $767.96 per million characters44 languagesMulti-agent systemMulti-agent system: median $56.78, from $0.4 to $1194.39 per million characters52 languagesEvery solution with a cost, at the list prices of August 2026, in every language where we evaluated it, so a range reflects the languages as well as themodels. Three-step solutions and multi-agent systems: the Intento LQA step is an estimate from the token counts of its requests. Fine-tuned models, the fifthlevel, are outside the report.
Cost of the solutions per million characters, by level of solution complexity
Key takeaway 40 · Costs

Meeting Standard requirements costs a similar amount to Basic in 47 of 53 languages, while Advanced costs more in 20.

We compare the cheapest solution that meets the requirements of each quality level in the same languages, those where two solutions meet every Advanced requirement. In most of them, the cheapest solution that meets general translation quality (no critical error, at most three major errors and at most 20 weighted errors per 1,000 words) also meets the other Standard requirements, or meets them as a two-step solution (translation + repair by code) at no extra cost.

In Marathi and Thai, only three-step solutions (translation + LQA + repair) meet the Standard requirements, at ~$550 per million characters against $8.20 and $21.28 for the Basic requirements.

Meeting the Advanced requirements costs 10 times as much or more in five languages, such as Estonian ($7.53 to $607 per million characters). We check gender neutrality only on the two best solutions of each language, those with the fewest errors, so the cheapest solution that meets the Advanced requirements often runs on a more expensive model or takes more steps.

Where the gender-neutrality agent works with a three-step solution, its gender checks add ~$400 per million characters, as in Amharic ($612 to $1,043).

See details →  ·  See more on Costs →

Languages by the cost change between sets of quality requirementsAbout the same (within 10%)10% more to 2 times as much2 to 10 times as much10 times as much or more0204060Basic to StandardBasic to Standard: About the same (within 10%) 47 (88.7%)47Basic to Standard: 10% more to 2 times as much 2 (3.8%)2Basic to Standard: 2 to 10 times as much 1 (1.9%)Basic to Standard: 10 times as much or more 3 (5.7%)3Standard to AdvancedStandard to Advanced: About the same (within 10%) 33 (62.3%)33Standard to Advanced: 10% more to 2 times as much 9 (17.0%)9Standard to Advanced: 2 to 10 times as much 6 (11.3%)6Standard to Advanced: 10 times as much or more 5 (9.4%)5languages53 languages where two solutions meet every Advanced requirement. Per language, the cheapest solution that meetsevery requirement of the level, at the list prices of August 2026. For Advanced, the cheaper of the two bestsolutions of the language that meets them. Three-step costs include an estimated LQA cost.
Languages by the cost change between sets of quality requirements
Key takeaway 41 · Performance

We compare translation time per 1,000 source words because the test texts differ in length.

The four source texts range from 270 to 999 words. We model translation time as a fixed time per request plus a time per word. With MT engines, most of it is the fixed time (~1.7 seconds per request for Yandex Cloud Translate v2), and with most LLMs it is the time per word (~0.06 seconds per word for Anthropic Claude Sonnet 5), so with an MT engine a short text takes more time per word.

We record the time of each text while we translate it, and for the models we call through the Intento platform, we take it from the platform logs, so these times are indicative.

See details →  ·  See more on Performance →

Key takeaway 42 · Performance

No assessed real-time one-step solution meets every Standard requirement.

We assessed real-time one-step solutions in 84 languages, from MT engines such as DeepL Next Gen, Translated Lara 2 and Microsoft Translator and from OpenAI gpt-oss-120b. Only two meet general translation quality (no critical error, at most three major errors and at most 20 weighted errors per 1,000 words), DeepL Next Gen in German and Romanian. 4 out of 198 meet both tag handling and formatting. A repair of tags and formatting by code does not add time, but it leaves general translation quality as it is.

The one-step solutions that meet every Standard requirement come from LLMs adjusted with context in the request. The fastest is Google Gemini 3.5 Flash in Indonesian at ~12 seconds per 1,000 source words, and across the 38 languages where such a solution exists, the fastest one takes a median of ~65 seconds per 1,000 words. Within a language, the LLM solutions that meet every Standard requirement take ~1.1 times as long as the other LLM solutions at the median, and longer in 27 of the 38 languages.

See details →  ·  See more on Performance →

One-step solutions by translation time and the Standard requirements they meetMT engineLLM-based MT engineLLMOpen-weight LLMReal-time cut, 2 seconds per text0.1 s1 s10 s100 s1,000 s10,000 sseconds per 1,000 source words over the four texts, log scaleOther solutionsMeet every requirementGoogle Cloud Advanced Translation API, Kurdish (Kurmanji): 0.13 sGoogle Cloud Advanced Translation API, Scottish Gaelic: 0.16 sMicrosoft Translator, Chinese Simplified: 0.17 sMicrosoft Translator, Danish: 0.19 sMicrosoft Translator, Italian: 0.19 sGoogle Cloud Advanced Translation API, Sinhala: 0.19 sMicrosoft Translator, Spanish (European): 0.2 sMicrosoft Translator, Estonian: 0.2 sMicrosoft Translator, Galician: 0.2 sMicrosoft Translator, Khmer: 0.2 sMicrosoft Translator, Urdu: 0.2 sMicrosoft Translator, Finnish: 0.21 sMicrosoft Translator, Portuguese (European): 0.21 sMicrosoft Translator, Telugu: 0.21 sMicrosoft Translator, Ukrainian: 0.21 sMicrosoft Translator, Afrikaans: 0.23 sMicrosoft Translator, German: 0.23 sMicrosoft Translator, Macedonian: 0.23 sMicrosoft Translator, Swedish: 0.23 sGoogle Cloud Advanced Translation API, Cebuano: 0.24 sMicrosoft Translator, Welsh: 0.24 sMicrosoft Translator, Lao: 0.24 sMicrosoft Translator, Croatian: 0.26 sMicrosoft Translator, Kannada: 0.26 sMicrosoft Translator, Norwegian: 0.26 sMicrosoft Translator, Albanian: 0.26 sMicrosoft Translator, Amharic: 0.27 sMicrosoft Translator, Catalan: 0.3 sModernMT Enterprise Edition, Finnish: 0.32 sMicrosoft Translator, Hindi: 0.32 sMicrosoft Translator, Romanian: 0.32 sMicrosoft Translator, Odia (Oriya): 0.34 sMicrosoft Translator, Greek: 0.35 sModernMT Enterprise Edition, Lithuanian: 0.35 sMicrosoft Translator, Malayalam: 0.36 sMicrosoft Translator, Hebrew: 0.38 sModernMT Enterprise Edition, Bosnian: 0.43 sMicrosoft Translator, Assamese: 0.46 sModernMT Enterprise Edition, Greek: 0.56 sMicrosoft Translator, Farsi: 0.56 sModernMT Enterprise Edition, Serbian: 1.01 sAmazon Translate, Dutch: 1.05 sAmazon Translate, Italian: 1.21 sAmazon Translate, German: 1.3 sAmazon Translate, French: 1.3 sAmazon Translate, Korean: 1.48 sAmazon Translate, Spanish (European): 1.52 sAmazon Translate, Japanese: 1.79 sElia Machine Translation System, Spanish (European): 2.96 sYandex Cloud Translate v2, Russian: 3.92 sYandex Cloud Translate v2, Belarusian: 4.75 sALL-Lab, Arabic (MSA): 35.23 sGoogle Translation LLM, Belarusian: 0.38 sGoogle Translation LLM, Slovak: 0.44 sGoogle Translation LLM, Polish: 0.46 sGoogle Translation LLM, Vietnamese: 0.48 sDeepL Next Gen, Malay: 0.5 sGoogle Translation LLM, Turkish: 0.5 sGoogle Translation LLM, Estonian: 0.51 sGoogle Translation LLM, Thai: 0.51 sDeepL Next Gen, Vietnamese: 0.52 sGoogle Translation LLM, Basque: 0.54 sGoogle Translation LLM, Latvian: 0.54 sGoogle Translation LLM, Azerbaijani: 0.55 sDeepL Next Gen, Czech: 0.55 sGoogle Translation LLM, Tamil: 0.55 sDeepL Next Gen, Galician: 0.56 sGoogle Translation LLM, Hungarian: 0.56 sGoogle Translation LLM, Nepali: 0.56 sDeepL Next Gen, Dutch: 0.56 sDeepL Next Gen, Sundanese: 0.56 sGoogle Translation LLM, Tagalog: 0.56 sDeepL Next Gen, Korean: 0.58 sDeepL Next Gen, Farsi: 0.59 sDeepL Next Gen, Croatian: 0.59 sDeepL Next Gen, Estonian: 0.6 sDeepL Next Gen, French: 0.6 sDeepL Next Gen, Polish: 0.6 sDeepL Next Gen, Swedish: 0.6 sGoogle Translation LLM, Swahili: 0.6 sDeepL Next Gen, Tagalog: 0.6 sDeepL Next Gen, Danish: 0.62 sDeepL Next Gen, Indonesian: 0.62 sDeepL Next Gen, Portuguese (European): 0.62 sGoogle Translation LLM, Armenian: 0.63 sGoogle Translation LLM, Kannada: 0.64 sDeepL Next Gen, Latvian: 0.64 sDeepL Next Gen, Slovak: 0.64 sGoogle Translation LLM, Czech: 0.66 sDeepL Next Gen, Slovenian: 0.66 sDeepL Next Gen, Afrikaans: 0.67 sDeepL Next Gen, Finnish: 0.67 sDeepL Next Gen, Icelandic: 0.67 sDeepL Next Gen, Albanian: 0.67 sDeepL Next Gen, Urdu: 0.67 sDeepL Next Gen, Macedonian: 0.69 sGoogle Translation LLM, Arabic (MSA): 0.7 sDeepL Next Gen, Azerbaijani: 0.7 sDeepL Next Gen, Welsh: 0.7 sDeepL Next Gen, German: 0.7 sGoogle Translation LLM, Icelandic: 0.7 sGoogle Translation LLM, Georgian: 0.7 sGoogle Translation LLM, Khmer: 0.7 sDeepL Next Gen, Uzbek: 0.7 sDeepL Next Gen, Tajik: 0.71 sGoogle Translation LLM, Irish: 0.73 sDeepL Next Gen, Armenian: 0.73 sGoogle Translation LLM, Lao: 0.73 sDeepL Next Gen, Turkish: 0.73 sDeepL Next Gen, Belarusian: 0.74 sDeepL Next Gen, Romanian: 0.74 sDeepL Next Gen, Tatar: 0.74 sGoogle Translation LLM, Bulgarian: 0.75 sDeepL Next Gen, Catalan: 0.75 sDeepL Next Gen, Nepali: 0.75 sDeepL Next Gen, Georgian: 0.77 sDeepL Next Gen, Tamil: 0.77 sDeepL Next Gen, Lithuanian: 0.78 sDeepL Next Gen, Bosnian: 0.81 sDeepL Next Gen, Hungarian: 0.81 sDeepL Next Gen, Telugu: 0.82 sDeepL Next Gen, Hausa: 0.83 sDeepL Next Gen, Ukrainian: 0.83 sDeepL Next Gen, Basque: 0.85 sDeepL Next Gen, Javanese: 0.85 sGoogle Translation LLM, Burmese: 0.85 sDeepL Next Gen, Arabic (MSA): 0.86 sDeepL Next Gen, Marathi: 0.86 sDeepL Next Gen, Irish: 0.87 sDeepL Next Gen, Bengali: 0.89 sGoogle Translation LLM, Tajik: 0.89 sGoogle Translation LLM, Punjabi: 0.9 sDeepL Next Gen, Hindi: 0.91 sGoogle Translation LLM, Mongolian: 0.91 sDeepL Next Gen, Hebrew: 0.94 sDeepL Next Gen, Chinese Simplified: 0.94 sDeepL Next Gen, Assamese: 1.01 sDeepL Next Gen, Japanese: 1.01 sDeepL Next Gen, Bulgarian: 1.02 sDeepL Next Gen, Gujarati: 1.06 sDeepL Next Gen, Russian: 1.07 sTranslated Lara 2, Danish: 1.09 sDeepL Next Gen, Punjabi: 1.1 sTranslated Lara 2, Indonesian: 1.13 sDeepL Next Gen, Italian: 1.14 sTranslated Lara 2, Swedish: 1.16 sDeepL Next Gen, Mongolian: 1.17 sTranslated Lara 2, Galician: 1.18 sDeepL Next Gen, Malayalam: 1.18 sTranslated Lara 2, Korean: 1.2 sTranslated Lara 2, Croatian: 1.24 sTranslated Lara 2, Slovenian: 1.25 sTranslated Lara 2, Bosnian: 1.26 sDeepL Next Gen, Thai: 1.26 sDeepL Next Gen, Greek: 1.3 sDeepL Next Gen, Pashto: 1.3 sTranslated Lara 2, Serbian: 1.3 sTranslated Lara 2, Russian: 1.32 sTranslated Lara 2, Afrikaans: 1.33 sTranslated Lara 2, Czech: 1.34 sTranslated Lara 2, Polish: 1.34 sTranslated Lara 2, Vietnamese: 1.34 sGoogle Translation LLM, Amharic: 1.36 sTranslated Lara 2, Turkish: 1.36 sTranslated Lara 2, Slovak: 1.37 sTranslated Lara 2, Romanian: 1.38 sTranslated Lara 2, Farsi: 1.42 sTranslated Lara 2, Ukrainian: 1.44 sTranslated Lara 2, Hungarian: 1.45 sTranslated Lara 2, Macedonian: 1.45 sGoogle Translation LLM, Kyrgyz: 1.48 sTranslated Lara 2, Albanian: 1.5 sTranslated Lara 2, Tagalog: 1.52 sTranslated Lara 2, Latvian: 1.58 sTranslated Lara 2, Irish: 1.61 sTranslated Lara 2, Urdu: 1.72 sDeepL Next Gen, Yiddish: 1.75 sTranslated Lara 2, Georgian: 1.8 sTranslated Lara 2, Thai: 1.8 sTranslated Lara 2, Belarusian: 1.84 sDeepL Next Gen, Norwegian: 1.84 sDeepL Next Gen, Chinese Traditional: 1.97 sDeepL Next Gen, Spanish (European): 2.0 sTranslated Lara 2, Hindi: 2.0 sTranslated Lara 2, Marathi: 2.12 sTranslated Lara 2, Tatar: 2.22 sTranslated Lara 2, Mongolian: 2.23 sTranslated Lara 2, Tajik: 2.26 sTranslated Lara 2, Kannada: 2.32 sTranslated Lara 2, Telugu: 2.38 sTranslated Lara 2, Tamil: 2.55 sTranslated Lara 2, Punjabi: 2.62 sTranslated Lara 2, Malayalam: 2.82 sTranslated Lara 2, Amharic: 5.16 sTranslated Lara 2, Lao: 5.45 sTranslated Lara 2, Khmer: 6.26 sTranslated Lara 2, Sinhala: 6.38 sGoogle Translation LLM, Marathi: 12.29 sWidn AI Tower Sugarloaf, Chinese Simplified: 35.8 sTranslated Lara 2, Odia (Oriya): 38.16 sWidn AI Tower Sugarloaf, Italian: 39.68 sWidn AI Tower Sugarloaf, Portuguese (European): 41.83 sWidn AI Tower Sugarloaf, Japanese: 42.24 sWidn AI Tower Sugarloaf, Spanish (European): 42.81 sWidn AI Tower Sugarloaf, Polish: 48.52 sWidn AI Tower Sugarloaf, Romanian: 49.48 sWidn AI Tower Sugarloaf, Hungarian: 51.2 sWidn AI Tower Sugarloaf, Ukrainian: 52.14 sWidn AI Tower Sugarloaf, Finnish: 60.69 sOpenAI gpt-oss-120b, Dutch: 1.36 sOpenAI gpt-oss-120b, French: 1.65 sOpenAI gpt-oss-120b, Indonesian: 1.85 sOpenAI gpt-oss-120b, Malay: 1.88 sOpenAI gpt-oss-120b, Arabic (MSA): 2.01 sOpenAI gpt-oss-120b, Catalan: 2.07 sOpenAI gpt-oss-120b, Pashto: 2.3 sOpenAI gpt-oss-120b, Norwegian: 2.39 sOpenAI gpt-oss-120b, Hebrew: 2.42 sOpenAI gpt-oss-120b, Welsh: 2.57 sOpenAI gpt-oss-120b, Bulgarian: 2.7 sOpenAI gpt-oss-120b, Tatar: 2.85 sMeta Llama 4 Maverick, Portuguese (European): 32.73 sMeta Llama 4 Maverick, German: 33.42 sSwiss AI Apertus-8B-2509, Chinese Simplified: 39.7 sAlibaba Qwen3-235B-A22B, Italian: 100.74 sAlibaba Qwen3-235B-A22B, Portuguese (European): 101.33 sGoogle Gemma 4, Norwegian: 130.11 sGoogle Gemma 4, Thai: 132.59 sGoogle Gemma 4, Vietnamese: 134.35 sGoogle Gemma 4, Bosnian: 138.24 sGoogle Gemma 4, Farsi: 143.85 sGoogle Gemma 4, Marathi: 147.67 sGoogle Gemma 4, Croatian: 148.02 sGoogle Gemma 4, Hindi: 148.41 sGoogle Gemma 4, Khmer: 154.43 sGoogle Gemma 4, Korean: 156.7 sGoogle Gemma 4, Afrikaans: 161.41 sGoogle Gemma 4, Bengali: 164.73 sGoogle Gemma 4, Russian: 166.43 sGoogle Gemma 4, Urdu: 169.21 sGoogle Gemma 4, Malayalam: 172.45 sGoogle Gemma 4, Gujarati: 177.25 sGoogle Gemma 4, Azerbaijani: 178.72 sGoogle Gemma 4, Amharic: 179.57 sGoogle Gemma 4, Assamese: 181.33 sGoogle Gemma 4, Punjabi: 182.23 sGoogle Gemma 4, Telugu: 185.17 sGoogle Gemma 4, Tagalog: 185.41 sGoogle Gemma 4, Swahili: 186.33 sGoogle Gemma 4, Sinhala: 187.91 sGoogle Gemma 4, Tamil: 192.53 sGoogle Gemma 4, Macedonian: 193.63 sGoogle Gemma 4, Kannada: 194.82 sGoogle Gemma 4, Albanian: 195.19 sGoogle Gemma 4, Arabic (MSA): 197.14 sGoogle Gemma 4, Turkish: 198.97 sGoogle Gemma 4, Italian: 199.5 sGoogle Gemma 4, Serbian: 202.47 sGoogle Gemma 4, Cebuano: 207.66 sGoogle Gemma 4, Polish: 237.47 sGoogle Gemma 4, Ukrainian: 250.26 sGoogle Gemma 4, Odia (Oriya): 284.97 sGoogle Gemma 4, Indonesian: 115.12 sOpenAI GPT-5.6 Luna, Portuguese (European): 11.48 sOpenAI GPT-5.6 Luna, German: 11.63 sOpenAI GPT-5.6 Luna, Italian: 12.09 sOpenAI GPT-5.6 Luna, French: 12.28 sGoogle Gemini 3.5 Flash, French: 12.89 sOpenAI GPT-5.6 Luna, Catalan: 13.35 sGoogle Gemini 3.5 Flash, Catalan: 13.45 sOpenAI GPT-4.1, Swedish: 13.46 sGoogle Gemini 3.5 Flash, Chinese Traditional: 13.49 sGoogle Gemini 3.5 Flash, Korean: 14.13 sGoogle Gemini 3.5 Flash, Hindi: 14.22 sGoogle Gemini 3.5 Flash, Serbian: 14.33 sOpenAI GPT-5.6 Luna, Sundanese: 14.43 sGoogle Gemini 3.5 Flash, Japanese: 14.48 sOpenAI GPT-5.6 Terra, Chinese Simplified: 14.49 sOpenAI GPT-5.6 Luna, Bengali: 14.71 sOpenAI GPT-4.1, Galician: 14.76 sOpenAI GPT-5.6 Luna, Georgian: 14.92 sGoogle Gemini 3.5 Flash, Bulgarian: 14.94 sGoogle Gemini 3.5 Flash, Croatian: 14.96 sGoogle Gemini 3.5 Flash, Bosnian: 14.98 sOpenAI GPT-5.6 Luna, Swahili: 15.07 sGoogle Gemini 3.5 Flash, Turkish: 15.25 sOpenAI GPT-4.1, Norwegian: 15.27 sOpenAI GPT-4.1, French: 15.35 sGoogle Gemini 3.5 Flash, Slovenian: 15.76 sGoogle Gemini 3.5 Flash, Polish: 15.78 sOpenAI GPT-5.6 Luna, Kazakh: 15.82 sOpenAI GPT-5.6 Luna, Farsi: 15.88 sOpenAI GPT-5.6 Luna, Uzbek: 15.93 sGoogle Gemini 3.5 Flash, Macedonian: 15.98 sOpenAI GPT-4.1, German: 16.0 sGoogle Gemini 3.5 Flash, Tagalog: 16.0 sGoogle Gemini 3.5 Flash, Gujarati: 16.05 sGoogle Gemini 3.5 Flash, Belarusian: 16.08 sxAI Grok 4.20 (non-reasoning), Turkish: 16.33 sGoogle Gemini 3.5 Flash, Estonian: 16.72 sOpenAI GPT-5.6 Luna, Irish: 16.79 sGoogle Gemini 3.5 Flash, Georgian: 16.88 sOpenAI GPT-4.1, Finnish: 16.92 sGoogle Gemini 3.5 Flash, Marathi: 17.02 sOpenAI GPT-4.1, Bosnian: 17.06 sGoogle Gemini 3.5 Flash, Swahili: 17.06 sOpenAI GPT-5.6 Terra, Korean: 17.09 sOpenAI GPT-5.6 Luna, Frisian: 17.14 sGoogle Gemini 3.5 Flash, Farsi: 17.17 sOpenAI GPT-5.6 Luna, Telugu: 17.27 sGoogle Gemini 3.5 Flash, Telugu: 17.41 sGoogle Gemini 3.5 Flash, Kannada: 17.52 sGoogle Gemini 3.5 Flash, Malayalam: 17.54 sOpenAI GPT-4.1, Estonian: 17.6 sGoogle Gemini 3.5 Flash, Tamil: 17.64 sOpenAI GPT-4.1, Arabic (MSA): 17.77 sGoogle Gemini 3.5 Flash, Latvian: 17.94 sGoogle Gemini 3.5 Flash, Frisian: 18.04 sGoogle Gemini 3.5 Flash, Greek: 18.09 sOpenAI GPT-5.6 Luna, Mongolian: 18.27 sOpenAI GPT-5.6 Luna, Hungarian: 18.29 sOpenAI GPT-5.6 Luna, Icelandic: 18.43 sGoogle Gemini 3.5 Flash, Azerbaijani: 18.5 sxAI Grok 4.20 (non-reasoning), Bosnian: 18.95 sOpenAI GPT-5.6 Luna, Welsh: 18.98 sGoogle Gemini 3.5 Flash, Uzbek: 19.02 sGoogle Gemini 3.5 Flash, Kazakh: 19.11 sOpenAI GPT-5.6 Luna, Assamese: 19.3 sOpenAI GPT-4.1, Czech: 19.5 sOpenAI GPT-5.6 Terra, Japanese: 19.56 sGoogle Gemini 3.5 Flash, Lithuanian: 19.74 sOpenAI GPT-4.1, Russian: 19.74 sGoogle Gemini 3.5 Flash, Pashto: 20.01 sOpenAI GPT-4.1, Turkish: 20.48 sxAI Grok 4.20 (non-reasoning), Frisian: 20.56 sOpenAI GPT-4.1, Malay: 20.6 sxAI Grok 4.20 (non-reasoning), Lithuanian: 21.06 sGoogle Gemini 3.5 Flash, Sesotho: 21.09 sOpenAI GPT-4.1, Latvian: 21.53 sOpenAI GPT-4.1, Danish: 21.64 sOpenAI GPT-4.1, Nepali: 21.73 sGoogle Gemini 3.5 Flash, Kyrgyz: 21.83 sOpenAI GPT-4.1, Swahili: 22.11 sOpenAI GPT-4.1, Ukrainian: 22.73 sOpenAI GPT-5.6 Terra, Polish: 22.86 sxAI Grok 4.20 (non-reasoning), Uzbek: 23.1 sGoogle Gemini 3.5 Flash, Lao: 23.28 sOpenAI GPT-5.6 Luna, Sinhala: 23.33 sGoogle Gemini 3.5 Flash, Tajik: 23.48 sOpenAI GPT-4.1, Armenian: 23.57 sOpenAI GPT-4.1, Catalan: 23.75 sOpenAI GPT-5.6 Terra, Swedish: 23.79 sOpenAI GPT-4.1, Greek: 23.9 sxAI Grok 4.20 (non-reasoning), Kurdish (Kurmanji): 23.99 sOpenAI GPT-4.1, Belarusian: 24.18 sOpenAI GPT-4.1, Hungarian: 24.55 sGoogle Gemini 3.5 Flash, Punjabi: 24.58 sGoogle Gemini 3.5 Flash, Burmese: 24.81 sOpenAI GPT-4.1, Marathi: 24.84 sOpenAI GPT-4.1, Slovenian: 25.48 sOpenAI GPT-4.1, Albanian: 25.8 sCohere Command A Translate, Chinese Simplified: 25.8 sOpenAI GPT-4.1, Dutch: 26.35 sOpenAI GPT-4.1, Farsi: 26.49 sxAI Grok 4.20 (non-reasoning), Kazakh: 26.74 sOpenAI GPT-5.6 Luna, Burmese: 26.81 sOpenAI GPT-4.1, Portuguese (European): 26.93 sCohere Command A Translate, Spanish (European): 27.07 sOpenAI GPT-4.1, Hindi: 27.56 sCohere Command A Translate, Portuguese (European): 27.94 sCohere Command A Translate, Indonesian: 28.17 sOpenAI GPT-4.1, Croatian: 28.18 sCohere Command A Translate, Italian: 28.23 sOpenAI GPT-5.6 Terra, Afrikaans: 28.46 sOpenAI GPT-5.6 Terra, Slovak: 28.54 sCohere Command A Translate, French: 28.91 sOpenAI GPT-4.1, Somali: 29.21 sCohere Command A Translate, Japanese: 29.52 sCohere Command A Translate, Korean: 29.72 sOpenAI GPT-5.6 Terra, Bosnian: 29.82 sCohere Command A Translate, German: 29.87 sxAI Grok 4.20 (non-reasoning), Kyrgyz: 30.21 sOpenAI GPT-4.1, Slovak: 30.25 sOpenAI GPT-4.1, Kannada: 30.38 sCohere Command A Translate, Vietnamese: 30.48 sOpenAI GPT-5.6 Terra, Albanian: 30.89 sMistral Medium 3.1, Russian: 31.44 sOpenAI GPT-5.6 Luna, Odia (Oriya): 31.78 sCohere Command A Translate, Arabic (MSA): 32.25 sCohere Command A Translate, Czech: 32.26 sMiniMax M2.5, Chinese Traditional: 32.36 sCohere Command A Translate, Galician: 32.89 sAnthropic Claude Haiku 4.5, Serbian: 33.16 sCohere Command A Translate, Hebrew: 33.23 sOpenAI GPT-5.6 Terra, Latvian: 33.53 sOpenAI GPT-5.6 Terra, Tagalog: 33.57 sOpenAI GPT-4.1, Romanian: 33.92 sOpenAI GPT-4.1, Basque: 34.33 sCohere Command A Translate, Malay: 34.44 sMistral Large 3, French: 35.5 sAnthropic Claude Haiku 4.5, Afrikaans: 35.66 sOpenAI GPT-4.1, Urdu: 35.7 sCohere Command A Translate, Greek: 36.12 sAnthropic Claude Sonnet 5, French: 36.23 sAnthropic Claude Sonnet 5, Indonesian: 36.76 sCohere Command A Translate, Dutch: 37.42 sCohere Command A Translate, Sundanese: 37.68 sOpenAI GPT-4.1, Bulgarian: 38.09 sOpenAI GPT-4.1, Sundanese: 38.29 sMistral Medium 3.1, Japanese: 38.62 sAnthropic Claude Sonnet 5, German: 38.64 sCohere Command A Translate, Romanian: 39.11 sMistral Medium 3.1, Polish: 39.48 sCohere Command A Translate, Ukrainian: 39.56 sOpenAI GPT-5.5, Italian: 39.64 sOpenAI GPT-5.6 Sol, Chinese Traditional: 39.83 sAnthropic Claude Sonnet 5, Russian: 39.88 sDeepSeek V3.2, German: 40.05 sOpenAI GPT-5.5, Chinese Traditional: 40.2 sAI21 Jamba Large 1.7, Spanish (European): 40.44 sxAI Grok 4.20 (reasoning), Galician: 40.94 sAnthropic Claude Sonnet 5, Swedish: 41.14 sxAI Grok 4.20 (non-reasoning), Yiddish: 41.77 sCohere Command A Translate, Catalan: 42.06 sAnthropic Claude Sonnet 5, Chinese Traditional: 42.16 sMistral Large 3, Russian: 42.28 sAnthropic Claude Sonnet 5, Dutch: 42.39 sOpenAI GPT-4.1, Icelandic: 42.43 sAnthropic Claude Sonnet 5, Japanese: 42.69 sAnthropic Claude Fable 5, Chinese Simplified: 42.89 sAnthropic Claude Opus 4.8, Spanish (European): 43.14 sDeepSeek V3.2, Danish: 43.2 sAnthropic Claude Sonnet 5, Korean: 43.38 sOpenAI GPT-5.6 Sol, Romanian: 43.41 sGoogle Gemini 3.5 Flash, Vietnamese: 43.47 sDeepSeek V3.2, Lithuanian: 43.51 sAnthropic Claude Fable 5, Portuguese (European): 43.91 sAnthropic Claude Sonnet 5, Norwegian: 44.22 sAnthropic Claude Fable 5, Russian: 44.24 sMistral Large 3, Chinese Simplified: 44.26 sxAI Grok 4.20 (reasoning), Hebrew: 45.22 sOpenAI GPT-5.5, Russian: 45.72 sAnthropic Claude Opus 4.8, French: 45.8 sAnthropic Claude Sonnet 5, Catalan: 46.18 sMistral Large 3, Dutch: 46.51 sOpenAI GPT-4.1, Wolof: 46.51 sOpenAI GPT-5.5, French: 47.04 sOpenAI GPT-4.1, Yiddish: 47.37 sxAI Grok 4.20 (reasoning), Catalan: 47.74 sAnthropic Claude Opus 4.8, Chinese Simplified: 48.13 sOpenAI GPT-4.1, Tajik: 48.23 sAnthropic Claude Sonnet 5, Polish: 48.33 sMistral Large 3, Swedish: 48.38 sAnthropic Claude Opus 4.8, German: 48.65 sAnthropic Claude Fable 5, Chinese Traditional: 48.84 sAnthropic Claude Fable 5, French: 48.97 sCohere Command A Translate, Serbian: 49.71 sAnthropic Claude Opus 4.8, Italian: 49.93 sMistral Large 3, Catalan: 49.97 sOpenAI GPT-5.6 Sol, Turkish: 50.19 sxAI Grok 4.20 (reasoning), Dutch: 50.56 sAnthropic Claude Sonnet 5, Arabic (MSA): 50.57 sMistral Large 3, Danish: 50.64 sAnthropic Claude Sonnet 4.6, Portuguese (European): 50.73 sAnthropic Claude Sonnet 5, Vietnamese: 50.8 sAnthropic Claude Sonnet 4.6, Italian: 51.2 sAnthropic Claude Fable 5, Japanese: 51.38 sOpenAI GPT-5.6 Terra, Kyrgyz: 51.38 sAI21 Jamba Large 1.7, Javanese: 51.47 sMistral Large 3, Vietnamese: 51.47 sxAI Grok 4.20 (reasoning), Kinyarwanda: 51.58 sMistral Large 3, Romanian: 51.67 sMistral Large 3, Malay: 52.37 sAnthropic Claude Fable 5, Korean: 52.67 sxAI Grok 4.20 (reasoning), Afrikaans: 52.91 sAnthropic Claude Opus 4.8, Japanese: 52.92 sMistral Large 3, Norwegian: 53.88 sAnthropic Claude Sonnet 5, Ukrainian: 53.93 sGoogle Gemini 3.5 Flash, Russian: 54.16 sAnthropic Claude Opus 4.8, Korean: 54.17 sOpenAI GPT-5.6 Sol, Ukrainian: 54.25 sAnthropic Claude Opus 4.8, Chinese Traditional: 54.55 sxAI Grok 4.20 (reasoning), Finnish: 54.82 sAnthropic Claude Sonnet 5, Bosnian: 55.18 sMistral Large 3, Chinese Traditional: 55.26 sMistral Large 3, Ukrainian: 55.35 sAnthropic Claude Opus 4.8, Galician: 55.65 sAnthropic Claude Sonnet 5, Serbian: 55.7 sMistral Large 3, Serbian: 55.81 sAnthropic Claude Sonnet 5, Hindi: 55.84 sAnthropic Claude Opus 4.8, Catalan: 56.83 sAnthropic Claude Sonnet 5, Czech: 56.88 sGoogle Gemini 3.5 Flash, Nepali: 56.92 sGoogle Gemini 3.5 Flash, Slovak: 57.5 sAnthropic Claude Sonnet 5, Afrikaans: 57.53 sDeepSeek V3.2, Swahili: 57.86 sGoogle Gemini 3.5 Flash, Urdu: 57.9 sAnthropic Claude Fable 5, Tagalog: 58.13 sGoogle Gemini 3.5 Flash, Bengali: 58.24 sOpenAI GPT-5.6 Sol, Thai: 58.31 sGoogle Gemini 3.5 Flash, Thai: 58.42 sAnthropic Claude Sonnet 4.6, Romanian: 58.51 sGoogle Gemini 3.5 Flash, Swedish: 58.68 sOpenAI GPT-4.1, Welsh: 59.15 sDeepSeek V3.2, Hungarian: 59.19 sOpenAI GPT-5.6 Sol, Serbian: 59.65 sMistral Large 3, Hebrew: 59.85 sAnthropic Claude Opus 4.8, Polish: 59.95 sAnthropic Claude Fable 5, Ukrainian: 60.03 sxAI Grok 4.20 (reasoning), Greek: 60.27 sAnthropic Claude Sonnet 4.6, Malay: 60.35 sGoogle Gemini 3.5 Flash, Basque: 60.4 sDeepSeek V3.2, Gujarati: 60.44 sAnthropic Claude Opus 4.8, Danish: 60.91 sAnthropic Claude Fable 5, Bosnian: 60.98 sxAI Grok 4.20 (reasoning), Spanish (European): 61.46 sGoogle Gemini 3.5 Flash, Hungarian: 61.46 sOpenAI GPT-5.5, Lithuanian: 61.54 sMistral Large 3, Lithuanian: 61.63 sAnthropic Claude Sonnet 4.6, Polish: 61.68 sGoogle Gemini 3.5 Flash, Arabic (MSA): 62.04 sAnthropic Claude Sonnet 5, Slovenian: 62.06 sAnthropic Claude Sonnet 4.6, Czech: 62.27 sMistral Large 3, Hindi: 62.47 sOpenAI GPT-5.5, Uzbek: 62.85 sAnthropic Claude Fable 5, Afrikaans: 62.89 sAnthropic Claude Opus 4.8, Turkish: 63.16 sMistral Large 3, Czech: 63.38 sOpenAI GPT-5.6 Terra, Armenian: 63.4 sAnthropic Claude Sonnet 5, Bengali: 63.53 sxAI Grok 4.20 (reasoning), Welsh: 63.61 sGoogle Gemini 3.5 Flash, Finnish: 63.61 sAnthropic Claude Sonnet 4.6, Serbian: 63.63 sAnthropic Claude Sonnet 4.6, Korean: 64.3 sDeepSeek V3.2, Burmese: 64.57 sGoogle Gemini 3.1 Pro Preview, Vietnamese: 64.61 sDeepSeek V3.2, Armenian: 64.62 sMistral Medium 3.1, Indonesian: 64.66 sDeepSeek V3.2, Slovenian: 64.82 sGoogle Gemini 3.1 Pro Preview, Chinese Traditional: 64.89 sGoogle Gemini 3.1 Pro Preview, Chinese Simplified: 65.32 sOpenAI GPT-5.5, Dutch: 65.33 sAnthropic Claude Fable 5, Hindi: 65.83 sAnthropic Claude Sonnet 4.6, Croatian: 66.0 sxAI Grok 4.20 (reasoning), Cebuano: 66.27 sAnthropic Claude Sonnet 5, Bulgarian: 66.49 sGoogle Gemini 3.5 Flash, Albanian: 66.55 sGoogle Gemini 3.5 Flash, Sinhala: 67.04 sAnthropic Claude Sonnet 4.6, Slovak: 67.23 sAnthropic Claude Sonnet 4.6, Slovenian: 67.79 sGoogle Gemini 3.5 Flash, Amharic: 68.11 sOpenAI GPT-5.5, Serbian: 68.3 sOpenAI GPT-5.5, Latvian: 68.41 sAnthropic Claude Fable 5, Swahili: 68.46 sAnthropic Claude Opus 4.8, Hindi: 68.8 sAnthropic Claude Sonnet 5, Estonian: 69.31 sDeepSeek V3.2, Odia (Oriya): 69.95 sGoogle Gemini 3.5 Flash, Afrikaans: 70.02 sAnthropic Claude Sonnet 4.6, Afrikaans: 70.13 sOpenAI GPT-5.5, Tagalog: 70.75 sMistral Large 3, Icelandic: 71.05 sAnthropic Claude Sonnet 4.6, Ukrainian: 71.38 sOpenAI GPT-5.6 Sol, Slovenian: 71.67 sGoogle Gemini 3.1 Pro Preview, Arabic (MSA): 72.18 sOpenAI GPT-5.5, Hungarian: 72.48 sAnthropic Claude Sonnet 4.6, Tagalog: 74.01 sAnthropic Claude Sonnet 4.6, Turkish: 74.32 sAnthropic Claude Fable 5, Latvian: 74.55 sAnthropic Claude Sonnet 5, Macedonian: 74.63 sAnthropic Claude Sonnet 4.6, Vietnamese: 74.63 sxAI Grok 4.20 (reasoning), Irish: 74.79 sGoogle Gemini 3.5 Flash, Odia (Oriya): 75.25 sDeepSeek V3.2, Malayalam: 75.3 sOpenAI GPT-5.5, Finnish: 75.62 sMistral Large 3, Tagalog: 75.66 sAnthropic Claude Sonnet 4.6, Bulgarian: 75.69 sGoogle Gemini 3.1 Pro Preview, Bengali: 75.73 sMistral Large 3, Bengali: 75.8 sAnthropic Claude Fable 5, Lithuanian: 76.16 sGoogle Gemini 3.1 Pro Preview, Tamil: 77.04 sGoogle Gemini 3.1 Pro Preview, Thai: 77.14 sAnthropic Claude Fable 5, Farsi: 77.46 sOpenAI GPT-5.6 Terra, Yiddish: 77.56 sDeepSeek V3.2, Lao: 77.61 sGoogle Gemini 3.1 Pro Preview, Tagalog: 77.66 sMistral Large 3, Thai: 78.27 sAnthropic Claude Sonnet 5, Marathi: 78.42 sAnthropic Claude Opus 4.8, Slovenian: 78.5 sGoogle Gemini 3.1 Pro Preview, Hindi: 78.55 sAnthropic Claude Sonnet 4.6, Sundanese: 78.75 sOpenAI GPT-5.6 Terra, Kannada: 79.03 sAnthropic Claude Sonnet 5, Urdu: 79.23 sAnthropic Claude Fable 5, Estonian: 79.4 sAnthropic Claude Sonnet 5, Albanian: 79.73 sGoogle Gemini 3.5 Flash, Welsh: 80.11 sMistral Large 3, Nepali: 80.12 sOpenAI GPT-5.5, Sundanese: 80.34 sOpenAI GPT-5.5, Farsi: 80.86 sGoogle Gemini 3.1 Pro Preview, Cebuano: 81.67 sGoogle Gemini 3.1 Pro Preview, Slovenian: 81.75 sOpenAI GPT-5.6 Sol, Hebrew: 81.8 sGoogle Gemini 3.1 Pro Preview, Italian: 81.8 sGoogle Gemini 3.1 Pro Preview, Sinhala: 81.91 sMistral Large 3, Marathi: 81.92 sAnthropic Claude Sonnet 4.6, Estonian: 82.03 sOpenAI GPT-5.5, Afrikaans: 82.14 sAnthropic Claude Opus 4.8, Bengali: 82.7 sGoogle Gemini 3.1 Pro Preview, Urdu: 83.95 sGoogle Gemini 3.1 Pro Preview, Lao: 84.08 sGoogle Gemini 3.1 Pro Preview, Hebrew: 84.11 sAnthropic Claude Fable 5, Cebuano: 84.38 sAnthropic Claude Sonnet 5, Welsh: 85.09 sMistral Large 3, Kannada: 85.37 sGoogle Gemini 3.1 Pro Preview, Swahili: 85.49 sGoogle Gemini 3.1 Pro Preview, Basque: 85.52 sAnthropic Claude Sonnet 5, Thai: 85.52 sGoogle Gemini 3.1 Pro Preview, Czech: 85.53 sOpenAI GPT-5.6 Sol, Azerbaijani: 85.76 sGoogle Gemini 3.1 Pro Preview, Kannada: 86.61 sAnthropic Claude Sonnet 4.6, Arabic (MSA): 86.62 sAnthropic Claude Sonnet 5, Nepali: 86.64 sMistral Large 3, Telugu: 86.8 sGoogle Gemini 3.1 Pro Preview, Danish: 86.82 sAnthropic Claude Sonnet 4.6, Latvian: 87.08 sAnthropic Claude Opus 4.8, Bulgarian: 87.12 sDeepSeek V3.2, Sinhala: 87.13 sOpenAI GPT-5.6 Sol, Basque: 87.24 sAnthropic Claude Fable 5, Uzbek: 87.41 sGoogle Gemini 3.5 Flash, Armenian: 87.67 sGoogle Gemini 3.1 Pro Preview, Nepali: 87.68 sGoogle Gemini 3.1 Pro Preview, Frisian: 87.95 sOpenAI GPT-5.5, Mongolian: 87.98 sDeepSeek V3.2, Frisian: 88.13 sAnthropic Claude Fable 5, Basque: 88.41 sGoogle Gemini 3.1 Pro Preview, Hungarian: 88.92 sGoogle Gemini 3.1 Pro Preview, Icelandic: 88.96 sAnthropic Claude Opus 4.8, Sundanese: 89.05 sOpenAI GPT-5.5, Basque: 89.19 sAnthropic Claude Fable 5, Azerbaijani: 89.58 sOpenAI GPT-5.5, Icelandic: 89.64 sGoogle Gemini 3.1 Pro Preview, Korean: 89.83 sAnthropic Claude Opus 4.8, Icelandic: 90.06 sGoogle Gemini 3.1 Pro Preview, Finnish: 90.56 sGoogle Gemini 3.1 Pro Preview, Romanian: 91.52 sMistral Large 3, Punjabi: 91.7 sGoogle Gemini 3.1 Pro Preview, Dutch: 92.13 sGoogle Gemini 3.1 Pro Preview, Tatar: 92.18 sGoogle Gemini 3.1 Pro Preview, Amharic: 92.55 sOpenAI GPT-5.5, Hausa: 92.89 sAnthropic Claude Fable 5, Marathi: 92.98 sGoogle Gemini 3.1 Pro Preview, French: 94.21 sxAI Grok 4.20 (reasoning), Albanian: 94.79 sAnthropic Claude Opus 4.8, Estonian: 95.3 sAnthropic Claude Opus 4.8, Hungarian: 95.47 sAnthropic Claude Opus 4.8, Latvian: 95.72 sAnthropic Claude Fable 5, Nepali: 95.81 sxAI Grok 4.20 (reasoning), Estonian: 95.94 sGoogle Gemini 3.1 Pro Preview, Telugu: 96.17 sOpenAI GPT-5.5, Cebuano: 96.27 sOpenAI GPT-5.5, Tajik: 96.49 sGoogle Gemini 3.1 Pro Preview, Welsh: 96.79 sGoogle Gemini 3.1 Pro Preview, Kyrgyz: 96.8 sAnthropic Claude Fable 5, Urdu: 96.95 sGoogle Gemini 3.1 Pro Preview, Marathi: 97.33 sGoogle Gemini 3.1 Pro Preview, Mongolian: 97.85 sGoogle Gemini 3.1 Pro Preview, Latvian: 98.23 sGoogle Gemini 3.1 Pro Preview, Japanese: 98.78 sGoogle Gemini 3.1 Pro Preview, Lithuanian: 99.23 sAnthropic Claude Sonnet 4.6, Hebrew: 100.01 sAnthropic Claude Opus 4.8, Frisian: 100.39 sOpenAI GPT-5.6 Sol, Hausa: 101.33 sAnthropic Claude Sonnet 4.6, Farsi: 101.37 sAnthropic Claude Sonnet 4.6, Urdu: 101.72 sGoogle Gemini 3.1 Pro Preview, Albanian: 102.63 sGoogle Gemini 3.1 Pro Preview, Kazakh: 103.14 sAnthropic Claude Fable 5, Thai: 103.32 sGoogle Gemini 3.1 Pro Preview, Georgian: 103.67 sGoogle Gemini 3.1 Pro Preview, Tajik: 103.85 sOpenAI GPT-5.5, Kyrgyz: 104.1 sAnthropic Claude Sonnet 4.6, Bengali: 104.2 sGoogle Gemini 3.1 Pro Preview, Gujarati: 104.38 sGoogle Gemini 3.1 Pro Preview, Assamese: 104.46 sAnthropic Claude Fable 5, Irish: 104.75 sAnthropic Claude Opus 4.8, Basque: 105.08 sGoogle Gemini 3.1 Pro Preview, Russian: 105.64 sGoogle Gemini 3.5 Flash, Irish: 105.78 sOpenAI GPT-5.5, Kazakh: 108.21 sAnthropic Claude Sonnet 4.6, Kazakh: 108.57 sAnthropic Claude Opus 4.8, Uzbek: 108.7 sAnthropic Claude Fable 5, Kazakh: 108.96 sOpenAI GPT-5.5, Tatar: 109.51 sGoogle Gemini 3.1 Pro Preview, Burmese: 111.12 sGoogle Gemini 3.1 Pro Preview, Estonian: 111.75 sOpenAI GPT-5.5, Pashto: 112.01 sOpenAI GPT-5.5, Shona: 112.26 sxAI Grok 4.20 (reasoning), Kannada: 112.34 sAnthropic Claude Opus 4.8, Kurdish (Kurmanji): 113.18 sGoogle Gemini 3.1 Pro Preview, Serbian: 113.5 sAnthropic Claude Sonnet 4.6, Belarusian: 114.25 sOpenAI GPT-5.5, Urdu: 114.36 sGoogle Gemini 3.1 Pro Preview, Irish: 115.29 sAnthropic Claude Fable 5, Assamese: 117.04 sAnthropic Claude Fable 5, Tajik: 118.33 sAnthropic Claude Opus 4.8, Thai: 119.18 sGoogle Gemini 3.1 Pro Preview, Malayalam: 119.31 sOpenAI GPT-5.5, Scottish Gaelic: 119.5 sGoogle Gemini 3.1 Pro Preview, Armenian: 119.84 sOpenAI GPT-5.6 Sol, Tajik: 119.95 sOpenAI GPT-5.5, Georgian: 120.16 sz.ai GLM-5, Chinese Simplified: 120.43 sOpenAI GPT-5.5, Marathi: 121.46 sOpenAI GPT-5.5, Welsh: 123.1 sGoogle Gemini 3.1 Pro Preview, Pashto: 123.87 sAnthropic Claude Sonnet 4.6, Tajik: 124.14 sMistral Large 3, Gujarati: 124.43 sAnthropic Claude Fable 5, Kyrgyz: 124.63 sAnthropic Claude Sonnet 4.6, Georgian: 125.14 sAnthropic Claude Opus 4.8, Irish: 126.26 sOpenAI GPT-5.5, Frisian: 127.84 sAnthropic Claude Fable 5, Tatar: 128.17 sOpenAI GPT-5.6 Sol, Urdu: 128.48 sGoogle Gemini 3.1 Pro Preview, Odia (Oriya): 128.76 sAnthropic Claude Opus 4.8, Kazakh: 129.58 sxAI Grok 4.20 (reasoning), Odia (Oriya): 130.58 sOpenAI GPT-5.6 Sol, Hindi: 131.27 sOpenAI GPT-5.6 Sol, Pashto: 131.28 sOpenAI GPT-5.5, Gujarati: 132.24 sAnthropic Claude Fable 5, Pashto: 134.21 sOpenAI GPT-5.6 Sol, Yiddish: 135.25 sAnthropic Claude Fable 5, Tamil: 135.39 sGoogle Gemini 3.1 Pro Preview, Punjabi: 135.46 sGoogle Gemini 3.1 Pro Preview, Uzbek: 135.55 sGoogle Gemini 3.1 Pro Preview, Scottish Gaelic: 135.78 sOpenAI GPT-5.6 Sol, Nepali: 137.66 sz.ai GLM-5, Chinese Traditional: 137.97 sAnthropic Claude Fable 5, Mongolian: 140.6 sAnthropic Claude Opus 4.8, Tajik: 142.93 sz.ai GLM-5, Galician: 145.48 sAnthropic Claude Opus 4.8, Assamese: 145.83 sOpenAI GPT-5.5, Malayalam: 146.02 sAnthropic Claude Sonnet 5, Telugu: 146.14 sAnthropic Claude Opus 4.8, Kyrgyz: 147.72 sAnthropic Claude Sonnet 5, Tamil: 148.15 sAnthropic Claude Opus 4.8, Tatar: 149.21 sz.ai GLM-5, Swedish: 157.27 sOpenAI GPT-5.5, Tamil: 159.1 sOpenAI GPT-5.5, Armenian: 159.13 sMoonshotAI Kimi K2.5, Slovak: 160.19 sOpenAI GPT-5.5, Kannada: 161.87 sAnthropic Claude Sonnet 4.6, Khmer: 162.89 sOpenAI GPT-5.5, Telugu: 164.75 sAnthropic Claude Opus 4.8, Finnish: 165.21 sAnthropic Claude Opus 4.8, Mongolian: 166.04 sMoonshotAI Kimi K2.5, Galician: 166.19 sOpenAI GPT-5.5, Yiddish: 166.72 sOpenAI GPT-5.5, Irish: 168.09 sAnthropic Claude Opus 4.8, Tamil: 170.53 sOpenAI GPT-5.6 Sol, Malayalam: 171.48 sGoogle Gemini 3.1 Pro Preview, Polish: 171.56 sz.ai GLM-5, Romanian: 171.61 sAnthropic Claude Sonnet 5, Amharic: 172.41 sOpenAI GPT-5.6 Sol, Gujarati: 173.62 sAnthropic Claude Fable 5, Yiddish: 173.85 sAnthropic Claude Opus 4.8, Telugu: 174.01 sOpenAI GPT-5.6 Sol, Marathi: 175.06 sOpenAI GPT-5.5, Assamese: 175.53 sAnthropic Claude Fable 5, Telugu: 175.53 sAnthropic Claude Sonnet 4.6, Lao: 176.44 sMoonshotAI Kimi K2.5, Chinese Traditional: 187.23 sz.ai GLM-5, Farsi: 194.43 sMoonshotAI Kimi K2.5, Dutch: 194.72 sAnthropic Claude Fable 5, Amharic: 199.84 sOpenAI GPT-5.6 Sol, Tamil: 200.07 sOpenAI GPT-5.6 Sol, Amharic: 201.4 sz.ai GLM-5, Macedonian: 206.34 sGoogle Gemini 3.1 Pro Preview, Azerbaijani: 206.35 sMoonshotAI Kimi K2.5, Bengali: 208.1 sz.ai GLM-5, Slovenian: 214.2 sOpenAI GPT-5.5, Amharic: 216.01 sMoonshotAI Kimi K2.5, Afrikaans: 216.33 sOpenAI GPT-5.6 Sol, Lao: 220.27 sz.ai GLM-5, Dutch: 220.5 sz.ai GLM-5, Slovak: 221.41 sOpenAI GPT-5.5, Sinhala: 221.67 sAnthropic Claude Opus 4.8, Amharic: 225.55 sOpenAI GPT-5.5, Punjabi: 229.83 sAnthropic Claude Sonnet 5, Gujarati: 231.75 sz.ai GLM-5, Croatian: 235.67 sMoonshotAI Kimi K2.5, Yoruba: 236.96 sMoonshotAI Kimi K2.5, Danish: 237.65 sAnthropic Claude Opus 4.8, Armenian: 243.05 sz.ai GLM-5, Frisian: 245.44 sAnthropic Claude Fable 5, Sinhala: 246.57 sMoonshotAI Kimi K2.5, Frisian: 247.54 sz.ai GLM-5, Hebrew: 252.63 sOpenAI GPT-5.6 Sol, Khmer: 253.34 sOpenAI GPT-5.6 Sol, Punjabi: 257.54 sMoonshotAI Kimi K2.5, Croatian: 257.61 sz.ai GLM-5, Basque: 264.92 sOpenAI GPT-5.5, Lao: 265.57 sz.ai GLM-5, Tagalog: 265.9 sz.ai GLM-5, Thai: 266.38 sz.ai GLM-5, Belarusian: 266.51 sz.ai GLM-5, Hindi: 267.6 sAnthropic Claude Opus 4.8, Georgian: 270.42 sMoonshotAI Kimi K2.5, Bulgarian: 270.46 sAnthropic Claude Opus 4.8, Sinhala: 274.96 sz.ai GLM-5, Pashto: 277.41 sOpenAI GPT-5.5, Burmese: 284.07 sz.ai GLM-5, Azerbaijani: 291.98 sz.ai GLM-5, Kazakh: 302.26 sAnthropic Claude Fable 5, Gujarati: 309.24 sAnthropic Claude Opus 4.8, Gujarati: 311.48 sz.ai GLM-5, Tatar: 337.14 sz.ai GLM-5, Bengali: 354.0 sAnthropic Claude Opus 4.8, Khmer: 361.16 sz.ai GLM-5, Khmer: 363.32 sMoonshotAI Kimi K2.5, Odia (Oriya): 376.08 sz.ai GLM-5, Armenian: 377.86 sOpenAI GPT-5.5, Odia (Oriya): 379.17 sz.ai GLM-5, Kannada: 421.69 sz.ai GLM-5, Lithuanian: 453.2 sz.ai GLM-5, Gujarati: 525.72 sz.ai GLM-5, Tamil: 537.49 sz.ai GLM-5, Punjabi: 765.27 sz.ai GLM-5, Malayalam: 779.38 sz.ai GLM-5, Lao: 3897.34 sGoogle Gemini 3.5 Flash, Indonesian: 11.94 sOpenAI GPT-5.6 Luna, Indonesian: 12.1 sOpenAI GPT-5.6 Luna, Norwegian: 12.14 sOpenAI GPT-5.6 Luna, Malay: 12.72 sGoogle Gemini 3.5 Flash, Malay: 13.49 sOpenAI GPT-5.6 Terra, Spanish (European): 17.45 sOpenAI GPT-4.1, Indonesian: 21.92 sOpenAI GPT-5.6 Terra, Danish: 24.67 sOpenAI GPT-5.6 Terra, Bulgarian: 25.99 sOpenAI GPT-5.6 Terra, Czech: 27.7 sOpenAI GPT-5.6 Terra, Macedonian: 31.66 sOpenAI GPT-5.6 Sol, Russian: 43.94 sOpenAI GPT-5.5, Turkish: 45.41 sOpenAI GPT-5.6 Sol, Galician: 47.71 sOpenAI GPT-5.5, Malay: 49.59 sOpenAI GPT-5.5, Norwegian: 50.18 sOpenAI GPT-5.5, Swedish: 50.83 sAnthropic Claude Fable 5, Norwegian: 52.56 sAnthropic Claude Fable 5, Galician: 52.59 sOpenAI GPT-5.5, Danish: 54.04 sOpenAI GPT-5.5, Slovenian: 54.95 sAnthropic Claude Fable 5, Turkish: 57.49 sAnthropic Claude Opus 4.8, Norwegian: 57.56 sAnthropic Claude Fable 5, Arabic (MSA): 57.8 sOpenAI GPT-5.6 Sol, Dutch: 58.07 sAnthropic Claude Fable 5, Romanian: 59.68 sGoogle Gemini 3.5 Flash, Galician: 59.88 sAnthropic Claude Fable 5, Croatian: 60.6 sAnthropic Claude Fable 5, Serbian: 62.04 sOpenAI GPT-5.6 Sol, Croatian: 62.26 sAnthropic Claude Fable 5, Slovenian: 63.01 sAnthropic Claude Fable 5, Czech: 63.13 sOpenAI GPT-5.5, Catalan: 63.34 sOpenAI GPT-5.5, Croatian: 64.93 sAnthropic Claude Fable 5, Slovak: 65.82 sOpenAI GPT-5.6 Sol, Finnish: 66.69 sAnthropic Claude Opus 4.8, Vietnamese: 66.82 sOpenAI GPT-5.5, Arabic (MSA): 67.15 sAnthropic Claude Fable 5, Vietnamese: 67.52 sGoogle Gemini 3.1 Pro Preview, Farsi: 68.09 sAnthropic Claude Fable 5, Bulgarian: 68.23 sAnthropic Claude Opus 4.8, Czech: 70.26 sGoogle Gemini 3.1 Pro Preview, Catalan: 71.54 sAnthropic Claude Fable 5, Bengali: 71.98 sGoogle Gemini 3.1 Pro Preview, Spanish (European): 72.79 sOpenAI GPT-5.5, Macedonian: 72.81 sAnthropic Claude Fable 5, Hungarian: 74.35 sOpenAI GPT-5.5, Azerbaijani: 74.4 sGoogle Gemini 3.1 Pro Preview, Indonesian: 75.58 sOpenAI GPT-5.6 Sol, Estonian: 78.84 sAnthropic Claude Fable 5, Macedonian: 79.11 sGoogle Gemini 3.1 Pro Preview, German: 79.3 sGoogle Gemini 3.1 Pro Preview, Bulgarian: 81.57 sGoogle Gemini 3.1 Pro Preview, Swedish: 82.3 sGoogle Gemini 3.1 Pro Preview, Ukrainian: 82.61 sGoogle Gemini 3.1 Pro Preview, Croatian: 84.47 sGoogle Gemini 3.1 Pro Preview, Norwegian: 84.85 sAnthropic Claude Fable 5, Albanian: 85.18 sAnthropic Claude Fable 5, Hebrew: 85.9 sGoogle Gemini 3.1 Pro Preview, Slovak: 87.02 sAnthropic Claude Fable 5, Welsh: 91.16 sAnthropic Claude Fable 5, Greek: 94.51 sAnthropic Claude Fable 5, Belarusian: 95.74 sGoogle Gemini 3.1 Pro Preview, Macedonian: 96.62 sGoogle Gemini 3.1 Pro Preview, Belarusian: 96.84 sGoogle Gemini 3.1 Pro Preview, Turkish: 97.86 sOpenAI GPT-5.5, Greek: 97.92 sGoogle Gemini 3.1 Pro Preview, Bosnian: 103.14 sOpenAI GPT-5.6 Sol, Belarusian: 113.77 sOpenAI GPT-5.5, Nepali: 118.31 sGoogle Gemini 3.1 Pro Preview, Greek: 138.74 sOpenAI GPT-5.6 Sol, Greek: 141.29 sOpenAI GPT-5.5, Bengali: 144.3 sz.ai GLM-5, Indonesian: 184.37 sOpenAI GPT-5.5, Khmer: 215.51 sAnthropic Claude Fable 5, Georgian: 259.15 sOne point per one-step solution and language, in 92 languages. Meet every requirement: every Standard requirement, general translation qualityincluded.Company, product and model names and logos are shown for informational purposes only. All trademarks and logos belong to their respectiveowners. Their use does not imply affiliation or endorsement.
One-step solutions by translation time and the Standard requirements they meet
Key takeaway 43 · Performance

Two-step solutions take ~twice as long as the same solutions in one step, and three-step solutions ~three times as long.

In each LLM step, we get the whole text again from the LLM, so a post-edit or a repair takes about as long as the translation. We estimate the LQA step at ~85 seconds per 1,000 words, with all its requests in parallel. Two-step solutions (translation + repair by code) take as long as the same solutions in one step, since the repair of tags and formatting by code does not add measurable time.

In the 42 languages where only two-step or three-step solutions meet every Standard requirement, the fastest of them takes a median of ~190 seconds per 1,000 words, against ~60 seconds per 1,000 words where a one-step solution does.

We implement these workflows in Intento Language Hub. Timings from platform logs cover the model calls made through the Hub. The Intento LQA duration here remains an estimate.

See details →  ·  See more on Performance →

Median time per 1,000 words of two-step and three-step solutions, by stepTranslationIntento LQA step (estimate)Post-edit or repair0100200300Two-step (translation + repair by code)Two-step (translation + repair by code): Translation 96 (100.0%)Two-step (translation + repair by code): Intento LQA step (estimate) 0 (0.0%)Two-step (translation + repair by code): Post-edit or repair 0 (0.0%)97 sTwo-step (translation + post-edit)Two-step (translation + post-edit): Translation 81 (49.3%)Two-step (translation + post-edit): Intento LQA step (estimate) 0 (0.0%)Two-step (translation + post-edit): Post-edit or repair 84 (50.7%)177 sThree-step (translation + LQA + repair)Three-step (translation + LQA + repair): Translation 83 (32.5%)Three-step (translation + LQA + repair): Intento LQA step (estimate) 86 (33.7%)Three-step (translation + LQA + repair): Post-edit or repair 86 (33.8%)260 sseconds per 1,000 source wordsMedians over the two-step and three-step solutions, in seconds per 1,000 source words over the four texts. Thetotal is the median of the solution totals. The LQA step is an estimate, with all its requests in parallel.
Median time per 1,000 words of two-step and three-step solutions, by step
Key takeaway 44 · Performance

Meeting Standard requirements takes a similar time to Basic in 45 of 53 languages, while Advanced takes longer in 20.

We compare the fastest solution that meets the requirements of each quality level in the same languages, those where two solutions meet every Advanced requirement. In German, the fastest solution that meets the Basic requirements is real-time (DeepL Next Gen, under 1 second per 1,000 words), and the fastest one that meets the Standard requirements takes 30 seconds per 1,000 words, a two-step solution (translation + post-edit) on OpenAI GPT-5.6 Luna.

Meeting the Advanced requirements takes 2 to 4.4 times as long in 10 languages, such as Bulgarian (26 to 116 seconds per 1,000 words). We check gender neutrality only on the two best solutions of each language, those with the fewest errors, so the fastest solution that meets the Advanced requirements often runs on a slower model or takes more steps.

See details →  ·  See more on Performance →

Languages by the translation-time change between sets of quality requirementsLess (by more than 10%)About the same (within 10%)10% more to 2 times as long2 to 10 times as long10 times as long or more0204060Basic to StandardBasic to Standard: Less (by more than 10%) 0 (0.0%)Basic to Standard: About the same (within 10%) 45 (84.9%)45Basic to Standard: 10% more to 2 times as long 4 (7.5%)4Basic to Standard: 2 to 10 times as long 3 (5.7%)3Basic to Standard: 10 times as long or more 1 (1.9%)Standard to AdvancedStandard to Advanced: Less (by more than 10%) 2 (3.8%)2Standard to Advanced: About the same (within 10%) 31 (58.5%)31Standard to Advanced: 10% more to 2 times as long 10 (18.9%)10Standard to Advanced: 2 to 10 times as long 10 (18.9%)10Standard to Advanced: 10 times as long or more 0 (0.0%)languages53 languages where two solutions meet every Advanced requirement. Per language, the fastest solution that meetsevery requirement of the level, in seconds per 1,000 words. For Advanced, the faster of the two best solutions ofthe language that meets them. We time a gender-neutral version on its own translations, so in Galician and Swedishit is faster than the solution it is built on. Times are indicative.
Languages by the translation-time change between sets of quality requirements

Ask a question

Read the report chapters in detail

01MethodHow the benchmark was run: source texts, languages, quality levels, automatic scoring, LLM-as-a-judge review and the selection of solutions. 02ModelsThe AI models and MT engines evaluated: their features, language coverage and results. 03HarnessesThe solutions built around the models: prompts, glossaries, post-editing and multi-step workflows, and what each adds. 04LanguagesResults per language: the quality level reached, the solutions that reach it, and the language tiers. 05QualityThe 10 requirements: how solutions meet them and where they fail. 06CostsPrices of the evaluated solutions and the cost of reaching each quality level per language. 07PerformanceLatency of the evaluated solutions and which of them reach quality in real time.

Automate translation to meet your requirements

See Intento Language Hub in action.

Book a demo
Leave feedback