The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Catalan against 10 requirements in three levels: the solutions that meet them, what they cost and how fast they are, from the 2026 benchmark of 57 AI models and MT engines.
A solution is a model plus its harness: prompts, glossaries and other context, pre- and post-processing. It meets a quality level when it meets every requirement of the level.
| Requirement | Description | Met when |
|---|---|---|
| Basic requirements for any company | ||
| General translation quality | Translation errors of every type, such as mistranslations, omissions and grammar errors, classified by type and severity following MQM, the industry standard. | No critical error, at most 3 major errors in the four test texts (3,034 words) and at most 20 errors per 1,000 words, weighted by severity (minor 1, major 5, critical 10). |
| Standard requirements adds the rules companies set for their content | ||
| Terminology | The approved translation of every glossary term, in any of its correct forms. | 100% of the glossary terms. We test it in the 16 languages with a glossary, so it does not apply in Catalan. |
| Tone of voice | The formal register (vostè) in every sentence in the second person or the imperative. | 100% of those sentences. |
| Tags | Every HTML tag of the source, with its attributes, in the translation. | 100% of the tags. |
| Unicode character integrity | Every character intact, with no corrupted characters and no letter from another script that only looks like the right one. | 100% of the characters. |
| Locale conventions | The punctuation marks of the target language in place of the English ones. | No pass-or-fail bar. We report it as a share, because whether to localize punctuation is a matter of house style. |
| HTML entities | No HTML character code, such as the code for a quotation mark, that the source does not carry. | No added HTML entity. |
| Valid JSON | The translated JSON document still parses. | The document parses. |
| Spacing at tag boundaries | The spaces next to HTML tags stay as in the source. | 100% of the tag boundaries. |
| Advanced requirements adds a requirement for some companies and use cases | ||
| Gender neutrality | A gender-neutral translation wherever the English source leaves the gender of a person open, as in “your manager” or “the expert”. | 100% of the sentences that refer to a person. |
Pick a quality level, or your own requirements, bars and constraints, to see the shortlisted solutions that meet them in Catalan, with a column for each requirement you select. Under Optimize for, choose which of them comes first: the one with the fewest total errors, the lowest cost or the shortest time. Click a column heading to sort by it.
The shortlisted solutions on 13 models, with their Intento LQA results. Total errors per 1,000 words add the weighted general translation quality errors (minor 1, major 5, critical 10) to the missed glossary terms, sentences in the wrong register, tag errors and sentences that fail gender neutrality, each counted once. Formatting counts as met or missed and does not add to the total errors. The models that did not make the shortlist have automatic scores only: the shortlist. Gender neutrality: the Advanced requirements count the two best solutions of the language, with the gender-neutrality agent where they need it. The results of the other solutions are shown for information.
The two solutions that meet every Standard requirement in Catalan differ in errors, cost and time. The one on Google Gemini 3.1 Pro Preview has 21.7 total errors per 1,000 words and costs $4.50 per million characters. The one on OpenAI GPT-5.5 has 25.9 total errors per 1,000 words and costs $10.40 per million characters, but takes 63 seconds per 1,000 words against 72. So Optimize for selects Google Gemini 3.1 Pro Preview for quality and cost, and OpenAI GPT-5.5 for latency.
For the Standard requirements, a cost cap of at most $10.39 per million characters leaves only the solution on Google Gemini 3.1 Pro Preview, and a time limit of at most 71 seconds per 1,000 words only the one on OpenAI GPT-5.5. None of the solutions that meet every Standard requirement fit a cap under $4.50 per million characters or a limit under 63 seconds per 1,000 words.
With the gender-neutrality agent, the solution on OpenAI GPT-5.5 meets every Advanced requirement at $11.22 per million characters, against $10.40 without it. It takes 68 seconds per 1,000 words against 63, and its total errors per 1,000 words fall from 25.9 to 13.5. With the agent, the solution on Google Gemini 3.1 Pro Preview costs less than the one on OpenAI GPT-5.5, $4.52 per million characters, but takes longer, 112 seconds per 1,000 words.
Google Gemini 3.1 Pro Preview
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Luna
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Luna
one-step LLM adjusted for Standard requirements
Google Gemini 3.1 Pro Preview
one-step LLM adjusted for Standard requirements
Google Gemini 3.1 Pro Preview
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
Google Gemini 3.1 Pro Preview
one-step with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
one-step with the gender-neutrality agent
OpenAI GPT-5.5
one-step with the gender-neutrality agent
Optimize for selects one of the solutions that meet the level: the one with the fewest total errors, the lowest cost or the shortest time. Where no solution meets the level, the card shows the available solution with the fewest total errors and what it misses. Total errors count every requirement with countable errors.
| Quality level | Tier | Solutions | Simplest solution | Model | Total errors per 1,000 words | Cost per million characters | Seconds per 1,000 words |
|---|---|---|---|---|---|---|---|
| Basic | Tier 1: simple solution + backup | ★×4 | one-step | OpenAI GPT-5.6 Luna | 29.5 | $0.41 | 13 |
| Standard | Tier 1: simple solution + backup | ★★ | one-step | Google Gemini 3.1 Pro Preview | 21.7 | $4.50 | 72 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | with the gender-neutrality agent | Google Gemini 3.1 Pro Preview | 9.2 | $4.52 | 112 |
One solution per level: the simplest that meets the level, the cheapest of them where several are equally simple. ★ a one-step solution, ☆ a solution with more steps or the gender-neutrality agent.
From the Basic to the Standard requirements, tone of voice and tag handling remove the solution on Anthropic Claude Opus 4.8, and tone of voice the one on OpenAI GPT-5.6 Luna. The Advanced requirements keep the other two, with the gender-neutrality agent. Under the Basic requirements, Optimize for selects Google Gemini 3.1 Pro Preview for quality and OpenAI GPT-5.6 Luna for cost and latency. Under the other two levels, it selects Google Gemini 3.1 Pro Preview for quality and cost and OpenAI GPT-5.5 for latency.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| OpenAI GPT-5.6 Luna | 0.41 | 18.8 | Misses a Standard requirement |
| Mistral Large 3 | 0.72 | 34.6 | Misses a Standard requirement |
| xAI Grok 4.20 (reasoning) | 1.12 | 25 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 2.92 | 34.3 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.36 | 28.3 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.5 | 11.9 | Meets every Standard requirement |
| Anthropic Claude Sonnet 5 | 5.41 | 24.4 | Misses a Standard requirement |
| Microsoft Translator | 10 | 58.3 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 10.4 | 16.8 | Meets every Standard requirement |
| Anthropic Claude Opus 4.8 | 13.6 | 12.2 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 31.6 | Misses a Standard requirement |
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 28.3 |
| Gender neutrality | 10.4 |
| Tone of voice | 2.3 |
| Tag handling | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
In Catalan, some one-step solutions miss tag handling or tone of voice by a small margin: translations from xAI Grok 4.20 (reasoning) and Cohere Command A Translate have 99.6% and 98.8% of the tags kept, and those from Mistral Large 3 have 97.9% of the sentences in the tone of voice.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Catalan | Class |
|---|---|---|
| Microsoft Translator | 58.3 | MT engines |
| OpenAI gpt-oss-120b | 46.8 | Open-weight LLMs |
| Cohere Command A Translate | 44.2 | LLMs |
| Mistral Large 3 | 34.6 | LLMs |
| OpenAI GPT-4.1 | 34.3 | LLMs |
| DeepL Next Gen | 31.6 | LLM-based MT engines |
| Google Gemini 3.5 Flash | 28.3 | LLMs |
| xAI Grok 4.20 (reasoning) | 25 | LLMs |
| Anthropic Claude Sonnet 5 | 24.4 | LLMs |
| OpenAI GPT-5.6 Luna | 18.8 | LLMs |
| OpenAI GPT-5.5 | 16.8 | LLMs |
| Anthropic Claude Opus 4.8 | 12.2 | LLMs |
| Google Gemini 3.1 Pro Preview | 11.9 | LLMs |
No translation into Catalan: Widn AI Tower Anthill (far too short, wrong language), Widn AI Tower Vesuvius (far too short, wrong language).
The one-step LLM solutions on the shortlist have 11.9 to 44.2 weighted errors per 1,000 words in Catalan, the LLM-based MT engine DeepL Next Gen 31.6, the open-weight OpenAI gpt-oss-120b 46.8 and the MT engine Microsoft Translator 58.3. All four solutions that meet every Basic requirement are on LLMs. All 11 shortlisted LLMs, the open-weight one included, support Catalan by our measured criteria. From Widn AI Tower Anthill and Widn AI Tower Vesuvius we got text far too short and in the wrong language, so we did not count it as a translation.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Anthropic Claude Fable 5 | 0.796 | 10 | 1.86 | 24 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.796 | 11 | 1.77 | 17 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.793 | 13 | 1.76 | 16 | No | Not shortlisted |
| DeepSeek V3.2 | 0.792 | 16 | 1.74 | 12 | No | Not shortlisted |
| z.ai GLM-5 | 0.792 | 17 | 1.75 | 14 | No | Not shortlisted |
| Google Gemma 4 | 0.792 | 18 | 1.71 | 9 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.790 | 20 | 1.79 | 19 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.789 | 21 | 1.84 | 21 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.785 | 22 | 1.98 | 26 | No | Not shortlisted |
| Translated Lara 2 | 0.784 | 23 | 2.11 | 28 | No | Not shortlisted |
| Pangeanic PangeaMT | 0.781 | 24 | 1.84 | 23 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.780 | 25 | 1.84 | 22 | No | Not shortlisted |
| Google Translation LLM | 0.779 | 26 | 2.11 | 29 | No | Not shortlisted |
| MiniMax M2.5 | 0.778 | 27 | 2.01 | 27 | No | Not shortlisted |
| Elia Machine Translation System | 0.777 | 28 | 1.89 | 25 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.777 | 29 | 1.77 | 18 | No | Not shortlisted |
| Amazon Translate | 0.769 | 30 | 2.29 | 30 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.762 | 31 | 2.60 | 32 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.761 | 32 | 2.42 | 31 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.721 | 33 | 3.20 | 34 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.711 | 34 | 3.17 | 33 | No | Not shortlisted |
| Cohere Tiny Aya | 0.612 | 35 | 6.18 | 35 | No | Not shortlisted |
| AppTek NMT | 0.558 | 36 | 7.27 | 36 | No | Not shortlisted |
| OpenAI GPT-5.6 Luna | 0.804 | 1 | 1.60 | 2 | Yes | Top 8 by both metrics |
| Cohere Command A Translate | 0.803 | 2 | 1.61 | 3 | Yes | Top 8 by both metrics |
| DeepL Next Gen | 0.802 | 3 | 1.58 | 1 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | 0.801 | 4 | 1.80 | 20 | Yes | Top 8 by COMET-Kiwi |
| xAI Grok 4.20 (reasoning) | 0.801 | 5 | 1.67 | 5 | Yes | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | 0.801 | 6 | 1.75 | 15 | Yes | Top 8 by COMET-Kiwi |
| Google Gemini 3.5 Flash | 0.799 | 7 | 1.73 | 10 | Yes | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.5 | 0.797 | 8 | 1.74 | 11 | Yes | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | 0.797 | 9 | 1.70 | 6 | Yes | Top 8 by MetricX |
| Mistral Large 3 | 0.793 | 12 | 1.71 | 7 | Yes | Top 8 by MetricX |
| Microsoft Translator | 0.792 | 14 | 1.64 | 4 | Yes | Top 8 by MetricX |
| OpenAI GPT-4.1 | 0.792 | 15 | 1.71 | 8 | Yes | Top 8 by MetricX |
| OpenAI gpt-oss-120b | 0.791 | 19 | 1.75 | 13 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| OpenAI GPT-5.6 Luna | Adjusted LLM | 1 | 0.804 | 2 | 1.60 | Top 8 by both metrics |
| Cohere Command A Translate | Adjusted LLM | 2 | 0.803 | 3 | 1.61 | Top 8 by both metrics |
| DeepL Next Gen | Off-the-shelf NMT | 3 | 0.802 | 1 | 1.58 | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 4 | 0.801 | 20 | 1.80 | Top 8 by COMET-Kiwi |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 5 | 0.801 | 5 | 1.67 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 6 | 0.801 | 15 | 1.75 | Top 8 by COMET-Kiwi |
| Google Gemini 3.5 Flash | Adjusted LLM | 7 | 0.799 | 10 | 1.73 | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.5 | Adjusted LLM | 8 | 0.797 | 11 | 1.74 | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 9 | 0.797 | 6 | 1.70 | Top 8 by MetricX |
| Anthropic Claude Fable 5 | Adjusted LLM | 10 | 0.796 | 24 | 1.86 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 11 | 0.796 | 17 | 1.77 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 12 | 0.793 | 7 | 1.71 | Top 8 by MetricX |
| Mistral Medium 3.1 | Adjusted LLM | 13 | 0.793 | 16 | 1.76 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 14 | 0.792 | 4 | 1.64 | Top 8 by MetricX |
| OpenAI GPT-4.1 | Adjusted LLM | 15 | 0.792 | 8 | 1.71 | Top 8 by MetricX |
| DeepSeek V3.2 | Adjusted LLM | 16 | 0.792 | 12 | 1.74 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 17 | 0.792 | 14 | 1.75 | Not shortlisted |
| Google Gemma 4 | Adjusted LLM | 18 | 0.792 | 9 | 1.71 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 19 | 0.791 | 13 | 1.75 | Real-time place |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 20 | 0.790 | 19 | 1.79 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 21 | 0.789 | 21 | 1.84 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 22 | 0.785 | 26 | 1.98 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 23 | 0.784 | 28 | 2.11 | Not shortlisted |
| Pangeanic PangeaMT | Off-the-shelf NMT | 24 | 0.781 | 23 | 1.84 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 25 | 0.780 | 22 | 1.84 | Not shortlisted |
| Google Translation LLM | Adjusted NMT | 26 | 0.779 | 29 | 2.11 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 27 | 0.778 | 27 | 2.01 | Not shortlisted |
| Elia Machine Translation System | Off-the-shelf NMT | 28 | 0.777 | 25 | 1.89 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 29 | 0.777 | 18 | 1.77 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 30 | 0.769 | 30 | 2.29 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 31 | 0.762 | 32 | 2.60 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 32 | 0.761 | 31 | 2.42 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 33 | 0.721 | 34 | 3.20 | Not shortlisted |
| SYSTRAN PNMT API | Off-the-shelf NMT | 34 | 0.711 | 33 | 3.17 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 35 | 0.612 | 35 | 6.18 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 36 | 0.558 | 36 | 7.27 | Not shortlisted |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.802 | — | 1.49 | Same model family as OpenAI GPT-5.6 Luna |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.803 | — | 1.67 | Same model family as OpenAI GPT-5.6 Luna |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.795 | — | 1.81 | Same model family as xAI Grok 4.20 (reasoning) |
Every model we got a translation from, as the shortlist compared it: off the shelf, without context in the request, or adjusted, with it. A model that did not make the shortlist is shown adjusted, unless off the shelf it scores higher by 0.03 or more on COMET-Kiwi. Ranks among 36 models, with each model family counted once, and the top 8 of each metric in bold. Higher is better on COMET-Kiwi, lower on MetricX. The scores rank models within this language and are not a requirement result.
COMET-Kiwi alone brought the solutions on Anthropic Claude Opus 4.8 and OpenAI GPT-5.5 onto the list, 15th and 11th on MetricX, and both meet general translation quality. Translations from Microsoft Translator, on the list through MetricX alone, have 58.3 weighted errors per 1,000 words, and those from Cohere Command A Translate, second on COMET-Kiwi, 44.2. Nearest the cut in the tail are Anthropic Claude Fable 5, at rank 10 on COMET-Kiwi, and Google Gemma 4, at rank 9 on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.6 Luna, one-step LLM adjusted for Standard requirements | $0.41 |
| Standard | Google Gemini 3.1 Pro Preview, one-step LLM adjusted for Standard requirements | $4.50 |
| Advanced | Google Gemini 3.1 Pro Preview, one-step with the gender-neutrality agent | $4.52 |
| Model | US dollars per million characters | Class |
|---|---|---|
| OpenAI GPT-5.6 Luna | $0.41 | LLMs |
| Mistral Large 3 | $0.72 | LLMs |
| xAI Grok 4.20 (reasoning) | $1.12 | LLMs |
| OpenAI GPT-4.1 | $2.92 | LLMs |
| Google Gemini 3.5 Flash | $3.36 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.5 | LLMs |
| Anthropic Claude Sonnet 5 | $5.41 | LLMs |
| Microsoft Translator | $10 | MT engines |
| OpenAI GPT-5.5 | $10.4 | LLMs |
| Anthropic Claude Opus 4.8 | $13.6 | LLMs |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Opus 4.8 | 0.42 | LLMs |
| Anthropic Claude Fable 5 | 0.419 | LLMs |
| Anthropic Claude Sonnet 5 | 0.419 | LLMs |
| Mistral Medium 3.1 | 0.354 | LLMs |
| Mistral Large 3 | 0.349 | LLMs |
| AI21 Jamba Large 1.7 | 0.331 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.323 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.323 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.317 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 0.316 | LLMs |
| Cohere Command A Translate | 0.308 | LLMs |
| z.ai GLM-5 | 0.308 | LLMs |
| DeepSeek V3.2 | 0.304 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.296 | LLMs |
| Google Gemini 3.5 Flash | 0.296 | LLMs |
| Google Gemma 4 | 0.294 | Open-weight LLMs |
| xAI Grok 4.20 (reasoning) | 0.293 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.29 | LLMs |
| OpenAI GPT-4.1 | 0.278 | LLMs |
| OpenAI GPT-5.5 | 0.276 | LLMs |
| OpenAI GPT-5.6 Terra | 0.276 | LLMs |
| OpenAI GPT-5.6 Luna | 0.275 | LLMs |
| OpenAI GPT-5.6 Sol | 0.275 | LLMs |
| Cohere Tiny Aya | 0.274 | LLMs |
| MiniMax M2.5 | 0.271 | LLMs |
With billing on tokens, the tokenizers produce a median 0.3 tokens per character of Catalan, 1.2 times their count for English. On the shortlist, one-step solutions on LLMs cost a median $3.36 per million characters, against $10.00 for Microsoft Translator and $27.50 for DeepL Next Gen. The cheapest solution that meets every Basic requirement costs $0.41 per million characters. The cheapest that meets every Advanced requirement, on Google Gemini 3.1 Pro Preview with the gender-neutrality agent, costs $4.52 per million characters.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | OpenAI GPT-5.6 Luna, one-step LLM adjusted for Standard requirements | 13 |
| Standard | OpenAI GPT-5.5, one-step LLM adjusted for Standard requirements | 63 |
| Advanced | OpenAI GPT-5.5, one-step with the gender-neutrality agent | 68 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Google Cloud Advanced Translation API | 0.2 | MT engines |
| Microsoft Translator | 0.3 | MT engines |
| ModernMT Enterprise Edition | 0.4 | MT engines |
| Google Translation LLM | 0.5 | LLM-based MT engines |
| Pangeanic PangeaMT | 0.6 | MT engines |
| SYSTRAN PNMT API | 0.7 | MT engines |
| DeepL Next Gen | 0.8 | LLM-based MT engines |
| Translated Lara 2 | 1.3 | LLM-based MT engines |
| Amazon Translate | 1.5 | MT engines |
| OpenAI gpt-oss-120b | 2.1 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 3.2 | MT engines |
| Elia Machine Translation System | 3.8 | MT engines |
| OpenAI GPT-5.6 Luna | 13 | LLMs |
| Google Gemini 3.5 Flash | 13 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 14 | LLMs |
| Cohere Tiny Aya | 17 | LLMs |
| OpenAI GPT-4.1 | 24 | LLMs |
| OpenAI GPT-5.6 Terra | 24 | LLMs |
| Anthropic Claude Haiku 4.5 | 29 | LLMs |
| Mistral Medium 3.1 | 35 | LLMs |
| Cohere Command A Translate | 42 | LLMs |
| MiniMax M2.5 | 43 | LLMs |
| Anthropic Claude Sonnet 5 | 46 | LLMs |
| DeepSeek V3.2 | 47 | LLMs |
| AI21 Jamba Large 1.7 | 47 | LLMs |
| xAI Grok 4.20 (reasoning) | 48 | LLMs |
| Anthropic Claude Fable 5 | 50 | LLMs |
| Mistral Large 3 | 50 | LLMs |
| OpenAI GPT-5.6 Sol | 56 | LLMs |
| Anthropic Claude Opus 4.8 | 57 | LLMs |
| Anthropic Claude Sonnet 4.6 | 58 | LLMs |
| OpenAI GPT-5.5 | 63 | LLMs |
| Google Gemini 3.1 Pro Preview | 72 | LLMs |
| Swiss AI Apertus-8B-2509 | 76 | Open-weight LLMs |
| Alibaba Qwen3-235B-A22B | 114 | Open-weight LLMs |
| AppTek NMT | 118 | MT engines |
| Google Gemma 4 | 187 | Open-weight LLMs |
| z.ai GLM-5 | 206 | LLMs |
| MoonshotAI Kimi K2.5 | 274 | LLMs |
Real-time models are those the shortlist counts as real-time, under 2 seconds per text across the languages we timed. The chart shows the time of each model into Catalan.
Nine of the ten real-time models in Catalan are MT engines or LLM-based MT engines. The other one is the open-weight OpenAI gpt-oss-120b, added to the shortlist as a real-time solution. The three real-time solutions on the shortlist have 31.6 to 58.3 weighted errors per 1,000 words, against 11.9 and 16.8 for the two solutions that meet every Standard requirement. The other one-step LLM solutions on the shortlist take from 13 to 72 seconds per 1,000 words.
| The two best solutions | Solution | Sentences about people that are gender-neutral | Meets every Advanced requirement | Cost per million characters | Seconds per 1,000 words |
|---|---|---|---|---|---|
| Google Gemini 3.1 Pro Preview | one-step, with the gender-neutrality agent | 100% | ✓ | $4.52 | 112 |
| OpenAI GPT-5.5 | one-step, with the gender-neutrality agent | 100% | ✓ | $11.22 | 68 |
For tone of voice in Catalan we check the formal register, with vostè as the form of address. For gender neutrality we check neutral nouns and rephrasing, paired forms only when unavoidable, no typographic forms such as · or @, and no default gender. With the gender-neutrality agent, both of the two best solutions have 100% of the sentences about people gender-neutral. Their weighted errors per 1,000 words fall from 11.9 to 9.2 on Google Gemini 3.1 Pro Preview and from 16.8 to 13.5 on OpenAI GPT-5.5.
The agreement of each pair of automatic metrics in Catalan is above its threshold, with no borderline case, and the three LQA judges agree on a median of 96% of the sentences of the 13 shortlisted solutions. So the shortlist and the Intento LQA results on this page rest on measurement we can rely on.
Leave feedback