The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into European Portuguese 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. |
| Tone of voice | The formal register (third-person polite forms) 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 European Portuguese, 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 11 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. Inconclusive: the three LQA judges agree on fewer than 90% of the sentences of the solution, so it has no general translation quality result (09-02). 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.
Two two-step solutions (translation + post-edit) meet every Standard requirement in European Portuguese, and neither leads on errors, cost and time at once. The one on Anthropic Claude Fable 5 has 17.6 total errors per 1,000 words against 24.1 for the one on OpenAI GPT-5.6 Luna. It costs $61.09 against $0.38 per million characters, and takes 85 against 12 seconds per 1,000 words. Optimizing for quality selects Anthropic Claude Fable 5, and optimizing for cost or latency selects OpenAI GPT-5.6 Luna.
A cost cap between $0.39 and $61.08 per million characters, or a time limit between 12 and 85 seconds per 1,000 words, leaves one solution for the Standard requirements, the two-step one (translation + post-edit) on OpenAI GPT-5.6 Luna, and none for the Advanced ones. With a limit of one step, two solutions meet the Basic requirements, the one-step ones on OpenAI GPT-5.6 Luna and Anthropic Claude Fable 5, at $0.38 and $27.74 per million characters, and none the Standard ones.
For the Advanced requirements, the gender-neutrality agent works with the two-step solution (translation + post-edit) on Anthropic Claude Fable 5. It raises the cost from $61.09 to $72.19 per million characters and the time from 85 to 104 seconds per 1,000 words, and the total errors fall from 17.6 to 14.2 per 1,000 words. The other of the two best solutions, on OpenAI GPT-5.6 Luna, costs $1,038 per million characters with the agent and takes 130 seconds per 1,000 words. It does not meet the Advanced requirements.
Anthropic Claude Sonnet 4.6
three-step
OpenAI GPT-5.6 Luna
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Luna
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
two-step
OpenAI GPT-5.6 Luna
two-step
OpenAI GPT-5.6 Luna
two-step
Anthropic Claude Fable 5
two-step with the gender-neutrality agent
Anthropic Claude Fable 5
two-step with the gender-neutrality agent
Anthropic Claude Fable 5
two-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 | ★★ ☆ | one-step | OpenAI GPT-5.6 Luna | 28.8 | $0.38 | 11 |
| Standard | Tier 2: two or more solutions, none simple | ☆☆ | two-step | OpenAI GPT-5.6 Luna | 24.1 | $0.38 | 12 |
| Advanced | Tier 3: one solution | ☆ | with the gender-neutrality agent | Anthropic Claude Fable 5 | 14.2 | $72.19 | 104 |
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.
The Standard requirements remove all three solutions that meet the Basic ones in European Portuguese: the three-step solution (translation + LQA + repair) on Anthropic Claude Sonnet 4.6 on terminology, with 92.6% of the glossary terms right, and the one-step solutions on Anthropic Claude Fable 5, on terminology too, and on OpenAI GPT-5.6 Luna, on tag handling. Under the Basic requirements, Optimize for selects the one on Anthropic Claude Sonnet 4.6 for quality and the one on OpenAI GPT-5.6 Luna for cost and latency.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| OpenAI GPT-5.6 Luna | 0.38 | 19.4 | Misses a Standard requirement |
| Alibaba Qwen3-235B-A22B | 1.06 | 64.6 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 2.66 | 40.2 | Misses a Standard requirement |
| Anthropic Claude Sonnet 4.6 | 6.11 | 29.3 | Misses a Standard requirement |
| Widn AI Tower Sugarloaf | 10 | 40.2 | Misses a Standard requirement |
| Microsoft Translator | 10 | 46.5 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 18.5 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 27.74 | 14.8 | Misses a Standard requirement |
Of the shortlisted one-step solutions in European Portuguese with a conclusive Intento LQA result, only two have fewer weighted errors than the one on OpenAI GPT-5.6 Luna, with 19.4 per 1,000 words at $0.38 per million characters, and both cost more. The one on the LLM-based MT engine DeepL Next Gen has 18.5 weighted errors per 1,000 words at $27.50 per million characters, and the one on Anthropic Claude Fable 5, an LLM, 14.8 at $27.74.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 34.8 |
| Gender neutrality | 3.1 |
| Tag handling | 1 |
| Tone of voice | 0.7 |
| Terminology | 0.4 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
In one step, translations from Cohere Command A Translate, Meta Llama 4 Maverick and OpenAI GPT-4.1 have 99.6% of the tags right, and those from Alibaba Qwen3-235B-A22B 99.2%, where we require 100%. Translations from DeepL Next Gen miss general translation quality on major errors alone, with 4 major errors and 18.5 weighted errors per 1,000 words.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in European Portuguese | Class |
|---|---|---|
| Alibaba Qwen3-235B-A22B | 64.6 | Open-weight LLMs |
| Microsoft Translator | 46.5 | MT engines |
| Meta Llama 4 Maverick | 46.1 | Open-weight LLMs |
| Widn AI Tower Sugarloaf | 40.2 | LLM-based MT engines |
| OpenAI GPT-4.1 | 40.2 | LLMs |
| Anthropic Claude Sonnet 4.6 | 29.3 | LLMs |
| Cohere Command A Translate | 28 | LLMs |
| OpenAI GPT-5.6 Luna | 19.4 | LLMs |
| DeepL Next Gen | 18.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | 14.8 | LLMs |
No translation into European Portuguese: Widn AI Tower Vesuvius (cut off part-way).
In one step in European Portuguese, translations from the five shortlisted LLMs that are not open-weight have 14.8 to 40.2 weighted errors per 1,000 words, from the two open-weight LLMs 46.1 and 64.6, from DeepL Next Gen and Widn AI Tower Sugarloaf, two of the LLM-based MT engines, 18.5 and 40.2, and from Microsoft Translator, the one MT engine, 46.5. Alibaba Qwen3-235B-A22B is the shortlisted LLM that does not meet our language-support criteria: translations from it have 64.6 weighted errors per 1,000 words, over the limit.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| z.ai GLM-5 | 0.804 | 10 | 1.46 | 12 | No | Not shortlisted |
| Google Gemini 3.1 Pro Preview | 0.804 | 11 | 1.49 | 17 | No | Not shortlisted |
| Anthropic Claude Sonnet 5 | 0.803 | 13 | 1.45 | 11 | No | Not shortlisted |
| OpenAI GPT-5.5 | 0.802 | 14 | 1.49 | 16 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.802 | 15 | 1.46 | 13 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.801 | 17 | 1.54 | 23 | No | Not shortlisted |
| Anthropic Claude Opus 4.8 | 0.801 | 18 | 1.45 | 10 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.801 | 19 | 1.54 | 21 | No | Not shortlisted |
| Mistral Large 3 | 0.800 | 20 | 1.48 | 15 | No | Not shortlisted |
| Pangeanic PangeaMT | 0.800 | 21 | 1.55 | 24 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.800 | 22 | 1.57 | 27 | No | Not shortlisted |
| Google Gemma 4 | 0.800 | 23 | 1.45 | 9 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.799 | 24 | 1.52 | 19 | No | Not shortlisted |
| DeepSeek V3.2 | 0.799 | 25 | 1.54 | 22 | No | Not shortlisted |
| Cohere Tiny Aya | 0.799 | 26 | 1.61 | 28 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.798 | 27 | 1.50 | 18 | No | Not shortlisted |
| Tilde TildeOpen LLM engine | 0.797 | 28 | 1.56 | 25 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.797 | 29 | 1.57 | 26 | No | Not shortlisted |
| Google Gemini 3.5 Flash | 0.796 | 30 | 1.62 | 29 | No | Not shortlisted |
| Google Translation LLM | 0.795 | 31 | 1.94 | 39 | No | Not shortlisted |
| MiniMax M2.5 | 0.794 | 32 | 1.75 | 34 | No | Not shortlisted |
| NiuTrans | 0.792 | 33 | 1.78 | 35 | No | Not shortlisted |
| SpeakLeash Bielik-11B-v3.0-Instruct | 0.792 | 34 | 1.84 | 37 | No | Not shortlisted |
| Amazon Translate | 0.790 | 35 | 1.73 | 32 | No | Not shortlisted |
| Supertext LLM | 0.790 | 36 | 1.71 | 31 | No | Not shortlisted |
| Baidu Translate API | 0.788 | 37 | 1.83 | 36 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.782 | 38 | 1.75 | 33 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.782 | 39 | 2.01 | 40 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.777 | 40 | 2.41 | 43 | No | Not shortlisted |
| Oracle Cloud Language Service | 0.777 | 41 | 1.90 | 38 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.769 | 42 | 2.23 | 41 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.756 | 43 | 2.27 | 42 | No | Not shortlisted |
| AppTek NMT | 0.692 | 44 | 3.60 | 44 | No | Not shortlisted |
| Widn AI Tower Anthill | 0.469 | 45 | 11.14 | 45 | No | Not shortlisted |
| DeepL Next Gen | 0.815 | 1 | 1.35 | 1 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Luna | 0.810 | 2 | 1.43 | 7 | Yes | Top 8 by both metrics |
| Widn AI Tower Sugarloaf | 0.808 | 3 | 1.39 | 2 | Yes | Top 8 by both metrics |
| Meta Llama 4 Maverick | 0.807 | 4 | 1.41 | 4 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.807 | 5 | 1.47 | 14 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 4.6 | 0.806 | 6 | 1.42 | 5 | Yes | Top 8 by both metrics |
| Cohere Command A Translate | 0.805 | 7 | 1.40 | 3 | Yes | Top 8 by both metrics |
| Alibaba Qwen3-235B-A22B | 0.805 | 8 | 1.42 | 6 | Yes | Top 8 by both metrics |
| Microsoft Translator | 0.804 | 9 | 1.52 | 20 | Yes | Real-time place |
| Translated Lara 2 | 0.803 | 12 | 1.63 | 30 | Yes | Real-time place |
| OpenAI GPT-4.1 | 0.802 | 16 | 1.44 | 8 | Yes | Top 8 by MetricX |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| DeepL Next Gen | Adjusted NMT | 1 | 0.815 | 1 | 1.35 | Top 8 by both metrics |
| OpenAI GPT-5.6 Luna | Adjusted LLM | 2 | 0.810 | 7 | 1.43 | Top 8 by both metrics |
| Widn AI Tower Sugarloaf | Adjusted NMT | 3 | 0.808 | 2 | 1.39 | Top 8 by both metrics |
| Meta Llama 4 Maverick | Adjusted LLM | 4 | 0.807 | 4 | 1.41 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 5 | 0.807 | 14 | 1.47 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 6 | 0.806 | 5 | 1.42 | Top 8 by both metrics |
| Cohere Command A Translate | Adjusted LLM | 7 | 0.805 | 3 | 1.40 | Top 8 by both metrics |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 8 | 0.805 | 6 | 1.42 | Top 8 by both metrics |
| Microsoft Translator | Off-the-shelf NMT | 9 | 0.804 | 20 | 1.52 | Real-time place |
| z.ai GLM-5 | Adjusted LLM | 10 | 0.804 | 12 | 1.46 | Not shortlisted |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 11 | 0.804 | 17 | 1.49 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 12 | 0.803 | 30 | 1.63 | Real-time place |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 13 | 0.803 | 11 | 1.45 | Not shortlisted |
| OpenAI GPT-5.5 | Adjusted LLM | 14 | 0.802 | 16 | 1.49 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 15 | 0.802 | 13 | 1.46 | Not shortlisted |
| OpenAI GPT-4.1 | Adjusted LLM | 16 | 0.802 | 8 | 1.44 | Top 8 by MetricX |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 17 | 0.801 | 23 | 1.54 | Not shortlisted |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 18 | 0.801 | 10 | 1.45 | Not shortlisted |
| ModernMT Enterprise Edition | Adjusted NMT | 19 | 0.801 | 21 | 1.54 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 20 | 0.800 | 15 | 1.48 | Not shortlisted |
| Pangeanic PangeaMT | Off-the-shelf NMT | 21 | 0.800 | 24 | 1.55 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 22 | 0.800 | 27 | 1.57 | Not shortlisted |
| Google Gemma 4 | Adjusted LLM | 23 | 0.800 | 9 | 1.45 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 24 | 0.799 | 19 | 1.52 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 25 | 0.799 | 22 | 1.54 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 26 | 0.799 | 28 | 1.61 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 27 | 0.798 | 18 | 1.50 | Not shortlisted |
| Tilde TildeOpen LLM engine | Adjusted NMT | 28 | 0.797 | 25 | 1.56 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 29 | 0.797 | 26 | 1.57 | Not shortlisted |
| Google Gemini 3.5 Flash | Adjusted LLM | 30 | 0.796 | 29 | 1.62 | Not shortlisted |
| Google Translation LLM | Off-the-shelf NMT | 31 | 0.795 | 39 | 1.94 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 32 | 0.794 | 34 | 1.75 | Not shortlisted |
| NiuTrans | Off-the-shelf NMT | 33 | 0.792 | 35 | 1.78 | Not shortlisted |
| SpeakLeash Bielik-11B-v3.0-Instruct | Off-the-shelf LLM | 34 | 0.792 | 37 | 1.84 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 35 | 0.790 | 32 | 1.73 | Not shortlisted |
| Supertext LLM | Off-the-shelf NMT | 36 | 0.790 | 31 | 1.71 | Not shortlisted |
| Baidu Translate API | Off-the-shelf NMT | 37 | 0.788 | 36 | 1.83 | Not shortlisted |
| SYSTRAN PNMT API | Adjusted NMT | 38 | 0.782 | 33 | 1.75 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Off-the-shelf LLM | 39 | 0.782 | 40 | 2.01 | Not shortlisted |
| Google Cloud Advanced Translation API | Adjusted NMT | 40 | 0.777 | 43 | 2.41 | Not shortlisted |
| Oracle Cloud Language Service | Off-the-shelf NMT | 41 | 0.777 | 38 | 1.90 | Not shortlisted |
| Yandex Cloud Translate v2 | Adjusted NMT | 42 | 0.769 | 41 | 2.23 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 43 | 0.756 | 42 | 2.27 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 44 | 0.692 | 44 | 3.60 | Not shortlisted |
| Widn AI Tower Anthill | Adjusted NMT | 45 | 0.469 | 45 | 11.14 | Not shortlisted |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.808 | — | 1.37 | Same model family as OpenAI GPT-5.6 Luna |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.807 | — | 1.41 | Same model family as OpenAI GPT-5.6 Luna |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.799 | — | 1.59 | 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 45 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.
Anthropic Claude Fable 5 is on the shortlist through COMET-Kiwi alone, fifth there and 14th on MetricX. Of the one-step solutions with a conclusive Intento LQA result, the one on that model has the fewest weighted errors, 14.8 per 1,000 words. OpenAI GPT-4.1 is on it through MetricX alone, eighth there and 16th on COMET-Kiwi, and translations from it have 40.2 weighted errors per 1,000 words and 1 critical error. Nearest the cut in the tail are Google Gemma 4, ninth on MetricX, and z.ai GLM-5, 10th on COMET-Kiwi.
| 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 | $0.38 |
| Standard | OpenAI GPT-5.6 Luna, two-step | $0.38 |
| Advanced | Anthropic Claude Fable 5, two-step with the gender-neutrality agent | $72.19 |
| Model | US dollars per million characters | Class |
|---|---|---|
| OpenAI GPT-5.6 Luna | $0.38 | LLMs |
| Alibaba Qwen3-235B-A22B | $1.06 | Open-weight LLMs |
| OpenAI GPT-4.1 | $2.66 | LLMs |
| Anthropic Claude Sonnet 4.6 | $6.11 | LLMs |
| Widn AI Tower Sugarloaf | $10 | LLM-based MT engines |
| Microsoft Translator | $10 | MT engines |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $27.74 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Opus 4.8 | 0.423 | LLMs |
| Anthropic Claude Sonnet 5 | 0.422 | LLMs |
| Anthropic Claude Fable 5 | 0.421 | LLMs |
| Mistral Medium 3.1 | 0.344 | LLMs |
| Mistral Large 3 | 0.338 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.312 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.31 | LLMs |
| MoonshotAI Kimi K2.5 | 0.3 | LLMs |
| Widn AI Tower Anthill | 0.284 | LLM-based MT engines |
| Alibaba Qwen3-235B-A22B | 0.283 | Open-weight LLMs |
| Widn AI Tower Sugarloaf | 0.283 | LLM-based MT engines |
| AI21 Jamba Large 1.7 | 0.28 | LLMs |
| DeepSeek V3.2 | 0.277 | LLMs |
| z.ai GLM-5 | 0.273 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.254 | LLMs |
| Google Gemini 3.5 Flash | 0.254 | LLMs |
| Google Gemma 4 | 0.254 | Open-weight LLMs |
| Cohere Command A Translate | 0.253 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.243 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.243 | LLMs |
| OpenAI GPT-4.1 | 0.243 | LLMs |
| OpenAI GPT-5.5 | 0.243 | LLMs |
| OpenAI GPT-5.6 Terra | 0.243 | LLMs |
| OpenAI GPT-5.6 Luna | 0.241 | LLMs |
| OpenAI GPT-5.6 Sol | 0.241 | LLMs |
| Cohere Tiny Aya | 0.237 | LLMs |
| MiniMax M2.5 | 0.233 | LLMs |
The tokenizer of Anthropic Claude Fable 5 produces 0.421 tokens per character of the translation into European Portuguese, against 0.241 for OpenAI GPT-5.6 Luna. On Anthropic Claude Fable 5, the cost rises from $27.74 per million characters in one step to $61.09 in two (translation + post-edit), and to $72.19 with the gender-neutrality agent.
| 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 | 11 |
| Standard | OpenAI GPT-5.6 Luna, two-step | 12 |
| Advanced | Anthropic Claude Fable 5, two-step with the gender-neutrality agent | 104 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.2 | MT engines |
| Google Cloud Advanced Translation API | 0.2 | MT engines |
| Google Translation LLM | 0.4 | LLM-based MT engines |
| DeepL Next Gen | 0.6 | LLM-based MT engines |
| ModernMT Enterprise Edition | 1 | MT engines |
| Translated Lara 2 | 1.1 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 1.2 | Open-weight LLMs |
| SYSTRAN PNMT API | 1.4 | MT engines |
| Pangeanic PangeaMT | 1.5 | MT engines |
| Amazon Translate | 2.4 | MT engines |
| Yandex Cloud Translate v2 | 2.6 | MT engines |
| Oracle Cloud Language Service | 3.5 | MT engines |
| Widn AI Tower Anthill | 7.6 | LLM-based MT engines |
| Tilde TildeOpen LLM engine | 9.6 | LLM-based MT engines |
| xAI Grok 4.20 (non-reasoning) | 10 | LLMs |
| OpenAI GPT-5.6 Luna | 11 | LLMs |
| Baidu Translate API | 11 | MT engines |
| Google Gemini 3.5 Flash | 13 | LLMs |
| Supertext LLM | 15 | LLM-based MT engines |
| Cohere Tiny Aya | 18 | LLMs |
| Alibaba Machine Translation API | 20 | MT engines |
| OpenAI GPT-5.6 Terra | 20 | LLMs |
| Anthropic Claude Haiku 4.5 | 24 | LLMs |
| OpenAI GPT-4.1 | 27 | LLMs |
| Cohere Command A Translate | 28 | LLMs |
| MiniMax M2.5 | 28 | LLMs |
| Mistral Medium 3.1 | 32 | LLMs |
| Meta Llama 4 Maverick | 33 | Open-weight LLMs |
| Anthropic Claude Sonnet 5 | 35 | LLMs |
| Swiss AI Apertus-8B-2509 | 35 | Open-weight LLMs |
| AI21 Jamba Large 1.7 | 38 | LLMs |
| Widn AI Tower Sugarloaf | 42 | LLM-based MT engines |
| Mistral Large 3 | 42 | LLMs |
| Anthropic Claude Fable 5 | 44 | LLMs |
| Anthropic Claude Opus 4.8 | 44 | LLMs |
| OpenAI GPT-5.6 Sol | 44 | LLMs |
| OpenAI GPT-5.5 | 46 | LLMs |
| xAI Grok 4.20 (reasoning) | 50 | LLMs |
| Anthropic Claude Sonnet 4.6 | 51 | LLMs |
| NiuTrans | 54 | MT engines |
| SpeakLeash Bielik-11B-v3.0-Instruct | 55 | Open-weight LLMs |
| DeepSeek V3.2 | 71 | LLMs |
| AppTek NMT | 95 | MT engines |
| Alibaba Qwen3-235B-A22B | 101 | Open-weight LLMs |
| Google Gemini 3.1 Pro Preview | 116 | LLMs |
| z.ai GLM-5 | 182 | LLMs |
| Google Gemma 4 | 205 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 258 | 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 European Portuguese.
The real-time models in European Portuguese are six MT engines, three LLM-based MT engines and the open-weight OpenAI gpt-oss-120b. Three of them are on the shortlist, DeepL Next Gen, Microsoft Translator and Translated Lara 2. None of the three meet formatting. Translations from Translated Lara 2 have no conclusive Intento LQA result on general translation quality.
| 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 |
|---|---|---|---|---|---|
| Anthropic Claude Fable 5 | two-step, with the gender-neutrality agent | 100% | ✓ | $72.19 | 104 |
| OpenAI GPT-5.6 Luna | three-step, with the gender-neutrality agent | 100% | ✗ | $1,038 | 130 |
For European Portuguese we check the formal register, with third-person polite forms, and gender-neutral wording through neutral nouns and rephrasing, with doubling only when unavoidable, no glyphs and no default gender. With the gender-neutrality agent, translations from both of the two best solutions have 100% of the sentences about people gender-neutral. The one on OpenAI GPT-5.6 Luna has 97.9% of the sentences in the formal register and 5 major errors, more than general translation quality allows.
For the 11 shortlisted solutions in European Portuguese, the median share of sentences on which the three LQA judges agree is 94.9%. Neither pair of automatic metrics is borderline: the lower bound of the agreement of each pair stays above its threshold. With agreement that high, the shortlist carries weight, and so do the Intento LQA error counts of the 10 shortlisted solutions where the three LQA judges agree on 90% or more of the sentences.
Leave feedback