The State of Translation Automation 2026

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 →

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 →

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 →

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 →

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 →

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 →

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 →

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 →

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 →

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 →

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 →

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 →

We seek a main solution and a backup that both meet the requirements. At least one is a simple, one-step solution in 54 of the 100 languages for Basic requirements, 38 for Standard and 13 for Advanced. The backup may need more steps.

At least two solutions qualify, but none is one-step, in another 25 languages for Basic, 40 for Standard and 40 for Advanced. In 18 languages, we cannot assess all the requirements. For each language and quality level, we also report errors, cost and time for the same qualifying solution, selected by quality, cost or latency in the chart settings.

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 →

We aim to identify at least two solutions that meet every requirement of a quality level: a main solution and a backup.

Tier 1: simple solution + backup. At least two solutions qualify, and at least one is one-step. This usually leaves scope to choose based on quality, cost or time.

Tier 2: two or more solutions, none simple. At least two qualify, but all need additional steps: translation + post-edit, translation + LQA + repair, or, for gender neutrality, the gender-neutrality agent. There may still be scope to optimize, but most options are slow, expensive or both.

Tier 3: one solution. Only one solution meets every requirement.

Tier 4: no solution. None of the assessed solutions meets every requirement, including the three-step and multi-agent solutions. In practice, complex bespoke workflows can achieve high quality in these languages, sometimes reaching 9+ steps for Standard requirements. These workflows are outside the scope of this report.

Tier 5: cannot assess. The automatic metrics or Intento LQA cannot assess quality.

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

The five high-resource languages are Chinese Simplified and Traditional, French, Japanese and Korean.

Languages in the same family often share a tier. All 12 Slavic languages are in tier 1 for Basic. The Uralic languages (Estonian, Finnish, Hungarian) are in tier 1 for Basic and Standard. The Dravidian languages (Kannada, Malayalam, Tamil, Telugu) are in tier 2 for Standard and Advanced.

All 18 languages we cannot assess are low-resource or medium-resource languages.

See more on Languages →

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 →

For Basic requirements, Afrikaans has one qualifying solution; Burmese and Sundanese have none.

For Standard requirements, Afrikaans and Frisian have one qualifying solution each; Burmese and Sundanese have none.

For Advanced requirements, 11 languages have no qualifying solution: Albanian, Arabic, Burmese, Czech, French, Hebrew, Polish, Sinhala, Sundanese, Urdu and Yiddish. The other 18 languages that fall short of two qualifying solutions have one each.

See details →  ·  See more on Quality →

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 →

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 →

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 →

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 →

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 →

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 →

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 →

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 40 of the 82 languages, 25 of them genderless languages, which have no gendered nouns or pronouns. In 42 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 57% of the cases and translations from MT engines in 78%.

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 →

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 →

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 →

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 four languages, such as Telugu ($29.87 to $929 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 →

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 →

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 →

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 eight 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 →