The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Polish 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 Polish. |
| Tone of voice | The formal register (Państwo, with verbs in the third person plural) 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 Polish, 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 12 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.
Two solutions meet every Standard requirement in Polish. The three-step solution (translation + LQA + repair) on Anthropic Claude Opus 4.8 has the fewest errors, 18.2 total errors per 1,000 words, and Optimize for selects it for quality. It costs $529 per million characters and takes 197 seconds per 1,000 words. For cost and latency Optimize for selects the two-step solution (translation + post-edit) on OpenAI GPT-5.6 Terra, at $10.51 per million characters and 47 seconds per 1,000 words, with 26.7 total errors per 1,000 words.
For the Standard requirements, a cost cap from $10.51 to $528 per million characters leaves only the two-step solution on OpenAI GPT-5.6 Terra. So do a time limit from 47 to 196 seconds per 1,000 words and a limit of two steps. None of the solutions meet every Standard requirement under a cap of at most $10.50 or a limit of at most 46 seconds, or in one step. Under the Basic requirements alone, the one-step solution on OpenAI GPT-5.6 Terra qualifies at $4.49 per million characters and 23 seconds per 1,000 words.
For the Advanced requirements, the gender-neutrality agent works with both solutions that meet every Standard requirement. On the three-step solution on Anthropic Claude Opus 4.8, total errors fall from 18.2 to 9.7 per 1,000 words, the cost rises from $529 to $900 per million characters, and the time barely changes. On the two-step solution on OpenAI GPT-5.6 Terra, total errors fall from 26.7 to 13.9 per 1,000 words, the cost rises to $11.82 per million characters and the time to 59 seconds per 1,000 words.
OpenAI GPT-5.6 Terra
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Terra
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Terra
one-step LLM adjusted for Standard requirements
Anthropic Claude Opus 4.8
three-step
OpenAI GPT-5.6 Terra
two-step
OpenAI GPT-5.6 Terra
two-step
Anthropic Claude Opus 4.8
three-step with the gender-neutrality agent
Misses tone of voice, gender neutrality.
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 Terra | 24.4 | $4.49 | 23 |
| Standard | Tier 2: two or more solutions, none simple | ☆☆ | two-step | OpenAI GPT-5.6 Terra | 26.7 | $10.51 | 47 |
| Advanced | Tier 4: no solution | No solution | — | — | — |
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.
Three solutions meet every Basic requirement in Polish, and the Standard requirements remove all three on tone of voice. The two one-step ones among them also miss tag handling. In their place come the two-step solution (translation + post-edit) on OpenAI GPT-5.6 Terra and the three-step solution (translation + LQA + repair) on Anthropic Claude Opus 4.8. Optimize for selects a solution on OpenAI GPT-5.6 Terra for quality, cost and latency under the Basic requirements, and for cost and latency under the Standard ones.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.13 | 43.5 | Misses a Standard requirement |
| Mistral Medium 3.1 | 1.03 | 65.3 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.32 | 36.3 | Misses a Standard requirement |
| OpenAI GPT-5.6 Terra | 4.49 | 13.5 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.51 | 25.4 | Misses a Standard requirement |
| Anthropic Claude Sonnet 5 | 6.07 | 35.3 | Misses a Standard requirement |
| Anthropic Claude Sonnet 4.6 | 6.76 | 35.6 | Misses a Standard requirement |
| Widn AI Tower Sugarloaf | 10 | 66.2 | Misses a Standard requirement |
| Anthropic Claude Opus 4.8 | 15.36 | 19.1 | Misses a Standard requirement |
| Google Translation LLM | 21.46 | 65.3 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 34.3 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 29.3 | Misses a Standard requirement |
Every shortlisted one-step solution in Polish that costs more than $4.49 per million characters has more weighted errors than the one on OpenAI GPT-5.6 Terra, 13.5 per 1,000 words. The four on LLM-based MT engines cost $10.00 to $27.50 per million characters and have 29.3 to 66.2 weighted errors per 1,000 words. Only three shortlisted ones cost less: those on Google Gemma 4, Mistral Medium 3.1 and Google Gemini 3.5 Flash, at $0.13, $1.03 and $3.32 per million characters. They have 43.5, 65.3 and 36.3 weighted errors per 1,000 words.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 35.5 |
| Gender neutrality | 9.9 |
| Tone of voice | 1.5 |
| Tag handling | 1.3 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
Two out of 12 one-step solutions in Polish meet general translation quality, fewer than meet tone of voice, tag handling or formatting. Neither of the two meets tone of voice. The five that meet tone of voice, all on LLMs, have 25.4 to 43.5 weighted errors per 1,000 words, over the bar. The two-step solution (translation + post-edit) on OpenAI GPT-5.6 Terra and the three-step solution (translation + LQA + repair) on Anthropic Claude Opus 4.8 meet tone of voice and tag handling. They have 17.5 and 8.6 weighted errors per 1,000 words.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Polish | Class |
|---|---|---|
| Widn AI Tower Sugarloaf | 66.2 | LLM-based MT engines |
| Mistral Medium 3.1 | 65.3 | LLMs |
| Google Translation LLM | 65.3 | LLM-based MT engines |
| Google Gemma 4 | 43.5 | Open-weight LLMs |
| Google Gemini 3.5 Flash | 36.3 | LLMs |
| Anthropic Claude Sonnet 4.6 | 35.6 | LLMs |
| Anthropic Claude Sonnet 5 | 35.3 | LLMs |
| Translated Lara 2 | 34.3 | LLM-based MT engines |
| DeepL Next Gen | 29.3 | LLM-based MT engines |
| Google Gemini 3.1 Pro Preview | 25.4 | LLMs |
| Anthropic Claude Opus 4.8 | 19.1 | LLMs |
| OpenAI GPT-5.6 Terra | 13.5 | LLMs |
No translation into Polish: Widn AI Tower Anthill (cut off part-way, wrong language), Widn AI Tower Vesuvius (cut off part-way).
In one step in Polish, translations from the seven shortlisted LLMs that are not open-weight have 13.5 to 65.3 weighted errors per 1,000 words, and those from the four LLM-based MT engines 29.3 to 66.2. No other MT engine made the shortlist. Translations from Google Translation LLM have 9 critical errors and those from Mistral Medium 3.1 have 1, the only critical errors among the shortlisted one-step solutions. Mistral Medium 3.1 is also the shortlisted LLM outside our language-support criteria.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| OpenAI GPT-5.5 | 0.786 | 9 | 1.49 | 9 | No | Not shortlisted |
| Cohere Command A Translate | 0.785 | 11 | 1.55 | 15 | No | Not shortlisted |
| OpenAI GPT-4.1 | 0.785 | 12 | 1.52 | 13 | No | Not shortlisted |
| Mistral Large 3 | 0.784 | 14 | 1.57 | 16 | No | Not shortlisted |
| Anthropic Claude Fable 5 | 0.783 | 15 | 1.50 | 10 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.782 | 16 | 1.75 | 26 | No | Not shortlisted |
| SpeakLeash Bielik-11B-v3.0-Instruct | 0.782 | 17 | 1.53 | 14 | No | Not shortlisted |
| xAI Grok 4.20 (non-reasoning) | 0.782 | 18 | 1.71 | 24 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.782 | 19 | 1.85 | 31 | No | Not shortlisted |
| DeepSeek V3.2 | 0.779 | 21 | 1.59 | 18 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.775 | 23 | 1.81 | 29 | No | Not shortlisted |
| Tilde TildeOpen LLM engine | 0.774 | 24 | 1.68 | 21 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.774 | 25 | 1.58 | 17 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.773 | 26 | 1.62 | 20 | No | Not shortlisted |
| Supertext LLM | 0.771 | 27 | 1.73 | 25 | No | Not shortlisted |
| MiniMax M2.5 | 0.770 | 28 | 1.79 | 28 | No | Not shortlisted |
| Microsoft Translator | 0.769 | 29 | 1.68 | 22 | No | Not shortlisted |
| z.ai GLM-5 | 0.767 | 30 | 1.82 | 30 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.765 | 31 | 2.22 | 34 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.761 | 32 | 1.79 | 27 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.759 | 33 | 1.91 | 32 | No | Not shortlisted |
| Pangeanic PangeaMT | 0.749 | 34 | 2.00 | 33 | No | Not shortlisted |
| Amazon Translate | 0.744 | 35 | 2.40 | 36 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.743 | 36 | 2.48 | 37 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.740 | 37 | 2.24 | 35 | No | Not shortlisted |
| Oracle Cloud Language Service | 0.727 | 38 | 2.61 | 40 | No | Not shortlisted |
| Baidu Translate API | 0.726 | 39 | 2.56 | 38 | No | Not shortlisted |
| Cohere Tiny Aya | 0.725 | 40 | 2.60 | 39 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.723 | 41 | 2.72 | 41 | No | Not shortlisted |
| AppTek NMT | 0.573 | 42 | 7.40 | 42 | No | Not shortlisted |
| DeepL Next Gen | 0.797 | 1 | 1.17 | 1 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.791 | 2 | 1.33 | 3 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Terra | 0.791 | 3 | 1.52 | 12 | Yes | Top 8 by COMET-Kiwi |
| Translated Lara 2 | 0.788 | 4 | 1.62 | 19 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | 0.787 | 5 | 1.46 | 7 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.786 | 6 | 1.50 | 11 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Opus 4.8 | 0.786 | 7 | 1.47 | 8 | Yes | Top 8 by both metrics |
| Widn AI Tower Sugarloaf | 0.786 | 8 | 1.33 | 2 | Yes | Top 8 by both metrics |
| Google Gemini 3.5 Flash | 0.785 | 10 | 1.45 | 5 | Yes | Top 8 by MetricX |
| Anthropic Claude Sonnet 4.6 | 0.784 | 13 | 1.42 | 4 | Yes | Top 8 by MetricX |
| Mistral Medium 3.1 | 0.780 | 20 | 1.45 | 6 | Yes | Top 8 by MetricX |
| Google Translation LLM | 0.777 | 22 | 1.70 | 23 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| DeepL Next Gen | Adjusted NMT | 1 | 0.797 | 1 | 1.17 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 2 | 0.791 | 3 | 1.33 | Top 8 by both metrics |
| OpenAI GPT-5.6 Terra | Adjusted LLM | 3 | 0.791 | 12 | 1.52 | Top 8 by COMET-Kiwi |
| Translated Lara 2 | Adjusted NMT | 4 | 0.788 | 19 | 1.62 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 5 | 0.787 | 7 | 1.46 | Top 8 by both metrics |
| Google Gemma 4 | Adjusted LLM | 6 | 0.786 | 11 | 1.50 | Top 8 by COMET-Kiwi |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 7 | 0.786 | 8 | 1.47 | Top 8 by both metrics |
| Widn AI Tower Sugarloaf | Adjusted NMT | 8 | 0.786 | 2 | 1.33 | Top 8 by both metrics |
| OpenAI GPT-5.5 | Adjusted LLM | 9 | 0.786 | 9 | 1.49 | Not shortlisted |
| Google Gemini 3.5 Flash | Adjusted LLM | 10 | 0.785 | 5 | 1.45 | Top 8 by MetricX |
| Cohere Command A Translate | Adjusted LLM | 11 | 0.785 | 15 | 1.55 | Not shortlisted |
| OpenAI GPT-4.1 | Adjusted LLM | 12 | 0.785 | 13 | 1.52 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 13 | 0.784 | 4 | 1.42 | Top 8 by MetricX |
| Mistral Large 3 | Adjusted LLM | 14 | 0.784 | 16 | 1.57 | Not shortlisted |
| Anthropic Claude Fable 5 | Adjusted LLM | 15 | 0.783 | 10 | 1.50 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 16 | 0.782 | 26 | 1.75 | Not shortlisted |
| SpeakLeash Bielik-11B-v3.0-Instruct | Adjusted LLM | 17 | 0.782 | 14 | 1.53 | Not shortlisted |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | 18 | 0.782 | 24 | 1.71 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 19 | 0.782 | 31 | 1.85 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 20 | 0.780 | 6 | 1.45 | Top 8 by MetricX |
| DeepSeek V3.2 | Adjusted LLM | 21 | 0.779 | 18 | 1.59 | Not shortlisted |
| Google Translation LLM | Adjusted NMT | 22 | 0.777 | 23 | 1.70 | Real-time place |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 23 | 0.775 | 29 | 1.81 | Not shortlisted |
| Tilde TildeOpen LLM engine | Off-the-shelf NMT | 24 | 0.774 | 21 | 1.68 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 25 | 0.774 | 17 | 1.58 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Off-the-shelf LLM | 26 | 0.773 | 20 | 1.62 | Not shortlisted |
| Supertext LLM | Off-the-shelf NMT | 27 | 0.771 | 25 | 1.73 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 28 | 0.770 | 28 | 1.79 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 29 | 0.769 | 22 | 1.68 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 30 | 0.767 | 30 | 1.82 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 31 | 0.765 | 34 | 2.22 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 32 | 0.761 | 27 | 1.79 | Not shortlisted |
| SYSTRAN PNMT API | Off-the-shelf NMT | 33 | 0.759 | 32 | 1.91 | Not shortlisted |
| Pangeanic PangeaMT | Off-the-shelf NMT | 34 | 0.749 | 33 | 2.00 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 35 | 0.744 | 36 | 2.40 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 36 | 0.743 | 37 | 2.48 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 37 | 0.740 | 35 | 2.24 | Not shortlisted |
| Oracle Cloud Language Service | Off-the-shelf NMT | 38 | 0.727 | 40 | 2.61 | Not shortlisted |
| Baidu Translate API | Off-the-shelf NMT | 39 | 0.726 | 38 | 2.56 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 40 | 0.725 | 39 | 2.60 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 41 | 0.723 | 41 | 2.72 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 42 | 0.573 | 42 | 7.40 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.789 | — | 1.43 | Same model family as OpenAI GPT-5.6 Terra |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.787 | — | 1.50 | Same model family as OpenAI GPT-5.6 Terra |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | — | 0.781 | — | 1.83 | Same model family as xAI Grok 4.20 (non-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 42 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.
The metrics and Intento LQA order the shortlist differently in Polish. Translations from DeepL Next Gen, first on both metrics, have 29.3 weighted errors per 1,000 words in one step. Those from OpenAI GPT-5.6 Terra, 12th on MetricX and on the list through COMET-Kiwi alone, have 13.5 weighted errors per 1,000 words. The three solutions on the list through MetricX alone all miss general translation quality. Nearest the cut is OpenAI GPT-5.5, 9th on both metrics.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.6 Terra, one-step LLM adjusted for Standard requirements | $4.49 |
| Standard | OpenAI GPT-5.6 Terra, two-step | $10.51 |
| Advanced | No solution | — |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.13 | Open-weight LLMs |
| Mistral Medium 3.1 | $1.03 | LLMs |
| Google Gemini 3.5 Flash | $3.32 | LLMs |
| OpenAI GPT-5.6 Terra | $4.49 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.51 | LLMs |
| Anthropic Claude Sonnet 5 | $6.07 | LLMs |
| Anthropic Claude Sonnet 4.6 | $6.76 | LLMs |
| Widn AI Tower Sugarloaf | $10 | LLM-based MT engines |
| Anthropic Claude Opus 4.8 | $15.36 | LLMs |
| Google Translation LLM | $21.46 | LLM-based MT engines |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Sonnet 5 | 0.484 | LLMs |
| Anthropic Claude Opus 4.8 | 0.483 | LLMs |
| Anthropic Claude Fable 5 | 0.482 | LLMs |
| AI21 Jamba Large 1.7 | 0.441 | LLMs |
| Mistral Medium 3.1 | 0.41 | LLMs |
| Mistral Large 3 | 0.406 | LLMs |
| MoonshotAI Kimi K2.5 | 0.384 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.361 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.359 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.332 | Open-weight LLMs |
| Widn AI Tower Sugarloaf | 0.33 | LLM-based MT engines |
| DeepSeek V3.2 | 0.324 | LLMs |
| z.ai GLM-5 | 0.318 | LLMs |
| MiniMax M2.5 | 0.307 | LLMs |
| OpenAI GPT-4.1 | 0.307 | LLMs |
| OpenAI GPT-5.5 | 0.306 | LLMs |
| OpenAI GPT-5.6 Sol | 0.306 | LLMs |
| OpenAI GPT-5.6 Terra | 0.306 | LLMs |
| OpenAI GPT-5.6 Luna | 0.305 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.295 | LLMs |
| Google Gemma 4 | 0.293 | Open-weight LLMs |
| Google Gemini 3.5 Flash | 0.292 | LLMs |
| Cohere Command A Translate | 0.288 | LLMs |
| Cohere Tiny Aya | 0.283 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.266 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.266 | LLMs |
At the median, a tokenizer produces 0.31 tokens per character of the Polish translation, 1.3 times as many as for the English source. The tokenizers of Anthropic Claude Opus 4.8 and Anthropic Claude Sonnet 5 produce 0.483 and 0.484 tokens per character, yet the one-step solutions on them cost $15.36 and $6.07 per million characters. On OpenAI GPT-5.6 Terra, the one-step solution costs $4.49 per million characters and the two-step one (translation + post-edit) $10.51, the cheapest for the Basic and for the Standard requirements.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | OpenAI GPT-5.6 Terra, one-step LLM adjusted for Standard requirements | 23 |
| Standard | OpenAI GPT-5.6 Terra, two-step | 47 |
| Advanced | No solution | — |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Google Cloud Advanced Translation API | 0.1 | MT engines |
| Microsoft Translator | 0.3 | MT engines |
| ModernMT Enterprise Edition | 0.4 | MT engines |
| Google Translation LLM | 0.5 | LLM-based MT engines |
| DeepL Next Gen | 0.6 | LLM-based MT engines |
| Pangeanic PangeaMT | 0.6 | MT engines |
| Amazon Translate | 1.2 | MT engines |
| Translated Lara 2 | 1.3 | LLM-based MT engines |
| SYSTRAN PNMT API | 1.7 | MT engines |
| Yandex Cloud Translate v2 | 2.1 | MT engines |
| OpenAI gpt-oss-120b | 2.6 | Open-weight LLMs |
| Oracle Cloud Language Service | 3.9 | MT engines |
| Tilde TildeOpen LLM engine | 8.4 | LLM-based MT engines |
| Baidu Translate API | 13 | MT engines |
| xAI Grok 4.20 (non-reasoning) | 14 | LLMs |
| OpenAI GPT-5.6 Luna | 14 | LLMs |
| OpenAI GPT-4.1 | 15 | LLMs |
| Supertext LLM | 15 | LLM-based MT engines |
| Google Gemini 3.5 Flash | 16 | LLMs |
| Alibaba Machine Translation API | 18 | MT engines |
| Cohere Tiny Aya | 20 | LLMs |
| OpenAI GPT-5.6 Terra | 23 | LLMs |
| Anthropic Claude Haiku 4.5 | 32 | LLMs |
| MiniMax M2.5 | 36 | LLMs |
| Cohere Command A Translate | 39 | LLMs |
| Mistral Medium 3.1 | 39 | LLMs |
| Swiss AI Apertus-8B-2509 | 41 | Open-weight LLMs |
| DeepSeek V3.2 | 42 | LLMs |
| Anthropic Claude Sonnet 5 | 48 | LLMs |
| Widn AI Tower Sugarloaf | 49 | LLM-based MT engines |
| xAI Grok 4.20 (reasoning) | 52 | LLMs |
| OpenAI GPT-5.5 | 53 | LLMs |
| Mistral Large 3 | 53 | LLMs |
| Anthropic Claude Fable 5 | 54 | LLMs |
| OpenAI GPT-5.6 Sol | 57 | LLMs |
| Anthropic Claude Opus 4.8 | 60 | LLMs |
| Anthropic Claude Sonnet 4.6 | 62 | LLMs |
| SpeakLeash Bielik-11B-v3.0-Instruct | 65 | Open-weight LLMs |
| AI21 Jamba Large 1.7 | 67 | LLMs |
| AppTek NMT | 95 | MT engines |
| Alibaba Qwen3-235B-A22B | 138 | Open-weight LLMs |
| Google Gemini 3.1 Pro Preview | 172 | LLMs |
| z.ai GLM-5 | 228 | LLMs |
| Google Gemma 4 | 237 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 259 | 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 Polish.
Of the 10 real-time models in Polish, six are MT engines, three LLM-based MT engines and one the open-weight OpenAI gpt-oss-120b. The shortlisted solutions on the three LLM-based MT engines, DeepL Next Gen, Translated Lara 2 and Google Translation LLM, miss general translation quality, tone of voice and formatting. The fastest solution that meets every Basic requirement takes 23 seconds per 1,000 words, and the fastest for the Standard ones 47.
| 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 Opus 4.8 | three-step, with the gender-neutrality agent | 96.2% | ✗ | $900 | 197 |
| OpenAI GPT-5.6 Terra | two-step, with the gender-neutrality agent | 98.1% | ✗ | $11.82 | 59 |
For Polish we check the formal register, with Państwo and verbs in the third person plural, and gender neutrality through neutral rephrasing plus impersonal forms, with no default gender. The gender-neutrality agent also changes tone of voice on one of the two best solutions. Without it, the three-step solution (translation + LQA + repair) on Anthropic Claude Opus 4.8 has the formal register in 100% of the sentences that call for it. With it, the share is 97.9%, so the solution misses tone of voice as well as gender neutrality.
Both pairs of automatic metrics clear their thresholds in Polish, and neither result is borderline. Across the 12 shortlisted solutions, the three LQA judges agree on a median of 95.8% of the sentences. The Intento LQA error counts on this page rest on that agreement. The shortlist rests on the ranking by COMET-Kiwi and MetricX.
Leave feedback