The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Russian 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 (вы (Вы for the reader)) 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 Russian, 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.
None of the five solutions that meet every Standard requirement leads on errors, cost and time at once. For quality, Optimize for selects the two-step solution (translation + post-edit) on Anthropic Claude Fable 5, with 15.4 total errors per 1,000 words at $59.73 per million characters and 88 seconds per 1,000 words. For cost it selects the one on Google Gemini 3.5 Flash, with 21.6 total errors per 1,000 words at $3.05 per million characters. For latency it selects OpenAI GPT-5.6 Sol, in one step, at 44 seconds per 1,000 words.
Constraints narrow the five quickly. A limit of one step leaves only the solution on OpenAI GPT-5.6 Sol. A cost cap from $3.06 to just under the $6.98 per million characters of OpenAI GPT-5.6 Sol leaves only the two-step solution on Google Gemini 3.5 Flash. A time limit from 44 to just under the 46 seconds per 1,000 words of OpenAI GPT-5.5 leaves only the one-step solution on OpenAI GPT-5.6 Sol. The two-step solution on Google Gemini 3.1 Pro Preview is the slowest of the five, at 196 seconds per 1,000 words.
For the Advanced requirements, the gender-neutrality agent on the two-step solution on Google Gemini 3.1 Pro Preview raises the cost from $27.34 to $28.91 per million characters and the time from 196 to 201 seconds per 1,000 words, and its total errors fall from 18.2 to 18.1 per 1,000 words. On the two-step solution on Anthropic Claude Fable 5, the agent lowers total errors from 15.4 to 11.5 per 1,000 words and raises the cost from $59.73 to $63.32 per million characters, but that solution does not meet every Advanced requirement.
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Google Gemini 3.5 Flash
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
Tied with Anthropic Claude Fable 5 (one-step LLM adjusted for Standard requirements).
Anthropic Claude Fable 5
two-step
Google Gemini 3.5 Flash
two-step
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
Google Gemini 3.1 Pro Preview
two-step with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
two-step with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
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 | ★×4 ☆ | one-step | Google Gemini 3.5 Flash | 21.6 | $3.05 | 54 |
| Standard | Tier 1: simple solution + backup | ★ ☆×4 | one-step | OpenAI GPT-5.6 Sol | 18.9 | $6.98 | 44 |
| Advanced | Tier 3: one solution | ☆ | with the gender-neutrality agent | Google Gemini 3.1 Pro Preview | 18.1 | $28.91 | 201 |
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 same five models have solutions that meet the Basic and the Standard requirements, four of them the Standard ones only as two-step solutions (translation + post-edit). Under both, Optimize for selects Anthropic Claude Fable 5 for quality, Google Gemini 3.5 Flash for cost and OpenAI GPT-5.6 Sol for latency, under the Basic ones tied with Anthropic Claude Fable 5 at 44 seconds per 1,000 words. Under the Advanced ones, it selects the only solution that meets them, on Google Gemini 3.1 Pro Preview with the gender-neutrality agent.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.13 | 24.7 | Misses a Standard requirement |
| Mistral Large 3 | 0.82 | 31.3 | Misses a Standard requirement |
| Mistral Medium 3.1 | 1 | 51.7 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 2.86 | 30.7 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.05 | 19.1 | Misses a Standard requirement |
| Yandex Cloud Translate v2 | 3.94 | 76.1 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.14 | 20.8 | Misses a Standard requirement |
| Anthropic Claude Sonnet 5 | 5.38 | 20.8 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 6.98 | 16.5 | Meets every Standard requirement |
| OpenAI GPT-5.5 | 10.1 | 17.8 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 38.2 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 27.06 | 12.9 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 21.4 | Misses a Standard requirement |
Of the solutions that meet every Standard requirement, a dearer one has fewer weighted errors, except the two-step one on OpenAI GPT-5.5, at $10.10 per million characters and 17.8 weighted errors per 1,000 words. The two-step solution on Google Gemini 3.5 Flash costs $3.05 per million characters and has 19.1 weighted errors per 1,000 words, the one-step one on OpenAI GPT-5.6 Sol $6.98 and 16.5, the two-step one on Google Gemini 3.1 Pro Preview $27.34 and 15.8, and the one on Anthropic Claude Fable 5 $59.73 and 12.9.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 21.4 |
| Gender neutrality | 2.5 |
| Terminology | 0.2 |
| Tone of voice | 0 |
| Tag handling | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
Translations from Anthropic Claude Sonnet 5, Mistral Large 3 and Yandex Cloud Translate v2 have 99.6%, 99.2% and 99.2% of the tags right, short of the bar of 100% for tag handling. On terminology, translations from the shortlisted LLMs and LLM-based MT engines have 97.1% to 100% of the glossary terms right, and those from the one MT engine, Yandex Cloud Translate v2, have 74.7%. The one-step solutions on Google Gemini 3.5 Flash and OpenAI GPT-5.5 miss formatting, and their two-step solutions (translation + post-edit) meet it.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Russian | Class |
|---|---|---|
| Yandex Cloud Translate v2 | 76.1 | MT engines |
| Mistral Medium 3.1 | 51.7 | LLMs |
| Translated Lara 2 | 38.2 | LLM-based MT engines |
| Mistral Large 3 | 31.3 | LLMs |
| OpenAI GPT-4.1 | 30.7 | LLMs |
| Google Gemma 4 | 24.7 | Open-weight LLMs |
| DeepL Next Gen | 21.4 | LLM-based MT engines |
| Google Gemini 3.1 Pro Preview | 20.8 | LLMs |
| Anthropic Claude Sonnet 5 | 20.8 | LLMs |
| Google Gemini 3.5 Flash | 19.1 | LLMs |
| OpenAI GPT-5.5 | 17.8 | LLMs |
| OpenAI GPT-5.6 Sol | 16.5 | LLMs |
| Anthropic Claude Fable 5 | 12.9 | LLMs |
No translation into Russian: Swiss AI Apertus-8B-2509 (cut off part-way, wrong language, wrong script), SpeakLeash Bielik-11B-v3.0-Instruct (wrong language, wrong script), Widn AI Tower Vesuvius (cut off part-way).
Translations from the shortlisted LLMs have 12.9 to 51.7 weighted errors per 1,000 words in Russian, and those from the open-weight Google Gemma 4 have 24.7, 1.91 times as many as translations from Anthropic Claude Fable 5. The shortlisted LLM that misses our language-support criteria is Mistral Medium 3.1, with two critical errors. We got no usable translation from three models: wrong language or script from Swiss AI Apertus-8B-2509 and SpeakLeash Bielik-11B-v3.0-Instruct, and text cut off part-way from Widn AI Tower Vesuvius.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Anthropic Claude Sonnet 4.6 | 0.810 | 11 | 1.11 | 12 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.809 | 13 | 1.16 | 15 | No | Not shortlisted |
| Cohere Command A Translate | 0.809 | 14 | 1.14 | 13 | No | Not shortlisted |
| Anthropic Claude Opus 4.8 | 0.808 | 15 | 1.15 | 14 | No | Not shortlisted |
| DeepSeek V3.2 | 0.807 | 17 | 1.19 | 19 | No | Not shortlisted |
| Widn AI Tower Sugarloaf | 0.806 | 18 | 1.17 | 16 | No | Not shortlisted |
| z.ai GLM-5 | 0.805 | 19 | 1.18 | 18 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.804 | 20 | 1.32 | 24 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.803 | 21 | 1.27 | 22 | No | Not shortlisted |
| Tilde TildeOpen LLM engine | 0.801 | 23 | 1.25 | 21 | No | Not shortlisted |
| MiniMax M2.5 | 0.800 | 24 | 1.23 | 20 | No | Not shortlisted |
| Google Translation LLM | 0.797 | 25 | 1.78 | 33 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.796 | 26 | 1.34 | 26 | No | Not shortlisted |
| Microsoft Translator | 0.795 | 27 | 1.38 | 29 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.795 | 28 | 1.31 | 23 | No | Not shortlisted |
| Amazon Translate | 0.789 | 29 | 1.35 | 27 | No | Not shortlisted |
| NiuTrans | 0.783 | 30 | 1.60 | 30 | No | Not shortlisted |
| Supertext LLM | 0.778 | 31 | 1.35 | 28 | No | Not shortlisted |
| Pangeanic PangeaMT | 0.775 | 32 | 1.69 | 31 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.774 | 33 | 1.83 | 34 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.770 | 34 | 1.75 | 32 | No | Not shortlisted |
| Baidu Translate API | 0.756 | 35 | 2.24 | 35 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.754 | 36 | 2.37 | 38 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.753 | 37 | 2.25 | 37 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.746 | 38 | 2.24 | 36 | No | Not shortlisted |
| Naver Papago NMT Commercial | 0.725 | 39 | 2.45 | 39 | No | Not shortlisted |
| Cohere Tiny Aya | 0.719 | 40 | 2.82 | 40 | No | Not shortlisted |
| AppTek NMT | 0.592 | 41 | 5.57 | 41 | No | Not shortlisted |
| TartuNLP | 0.319 | 42 | 14.04 | 42 | No | Not shortlisted |
| DeepL Next Gen | 0.818 | 1 | 0.98 | 1 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.816 | 2 | 1.08 | 9 | Yes | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | 0.815 | 3 | 1.09 | 10 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | 0.814 | 4 | 1.09 | 11 | Yes | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.6 Sol | 0.814 | 5 | 1.00 | 3 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.813 | 6 | 1.06 | 4 | Yes | Top 8 by both metrics |
| Yandex Cloud Translate v2 | 0.813 | 7 | 0.99 | 2 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.5 | 0.812 | 8 | 1.18 | 17 | Yes | Top 8 by COMET-Kiwi |
| Mistral Large 3 | 0.812 | 9 | 1.07 | 8 | Yes | Top 8 by MetricX |
| OpenAI GPT-4.1 | 0.811 | 10 | 1.06 | 5 | Yes | Top 8 by MetricX |
| Google Gemini 3.5 Flash | 0.810 | 12 | 1.07 | 6 | Yes | Top 8 by MetricX |
| Mistral Medium 3.1 | 0.808 | 16 | 1.07 | 7 | Yes | Top 8 by MetricX |
| Translated Lara 2 | 0.802 | 22 | 1.33 | 25 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| DeepL Next Gen | Adjusted NMT | 1 | 0.818 | 1 | 0.98 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 2 | 0.816 | 9 | 1.08 | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 3 | 0.815 | 10 | 1.09 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 4 | 0.814 | 11 | 1.09 | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.6 Sol | Adjusted LLM | 5 | 0.814 | 3 | 1.00 | Top 8 by both metrics |
| Google Gemma 4 | Adjusted LLM | 6 | 0.813 | 4 | 1.06 | Top 8 by both metrics |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 7 | 0.813 | 2 | 0.99 | Top 8 by both metrics |
| OpenAI GPT-5.5 | Adjusted LLM | 8 | 0.812 | 17 | 1.18 | Top 8 by COMET-Kiwi |
| Mistral Large 3 | Adjusted LLM | 9 | 0.812 | 8 | 1.07 | Top 8 by MetricX |
| OpenAI GPT-4.1 | Adjusted LLM | 10 | 0.811 | 5 | 1.06 | Top 8 by MetricX |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 11 | 0.810 | 12 | 1.11 | Not shortlisted |
| Google Gemini 3.5 Flash | Adjusted LLM | 12 | 0.810 | 6 | 1.07 | Top 8 by MetricX |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 13 | 0.809 | 15 | 1.16 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 14 | 0.809 | 13 | 1.14 | Not shortlisted |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 15 | 0.808 | 14 | 1.15 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 16 | 0.808 | 7 | 1.07 | Top 8 by MetricX |
| DeepSeek V3.2 | Adjusted LLM | 17 | 0.807 | 19 | 1.19 | Not shortlisted |
| Widn AI Tower Sugarloaf | Adjusted NMT | 18 | 0.806 | 16 | 1.17 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 19 | 0.805 | 18 | 1.18 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 20 | 0.804 | 24 | 1.32 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 21 | 0.803 | 22 | 1.27 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 22 | 0.802 | 25 | 1.33 | Real-time place |
| Tilde TildeOpen LLM engine | Adjusted NMT | 23 | 0.801 | 21 | 1.25 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 24 | 0.800 | 20 | 1.23 | Not shortlisted |
| Google Translation LLM | Off-the-shelf NMT | 25 | 0.797 | 33 | 1.78 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 26 | 0.796 | 26 | 1.34 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 27 | 0.795 | 29 | 1.38 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 28 | 0.795 | 23 | 1.31 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 29 | 0.789 | 27 | 1.35 | Not shortlisted |
| NiuTrans | Off-the-shelf NMT | 30 | 0.783 | 30 | 1.60 | Not shortlisted |
| Supertext LLM | Off-the-shelf NMT | 31 | 0.778 | 28 | 1.35 | Not shortlisted |
| Pangeanic PangeaMT | Off-the-shelf NMT | 32 | 0.775 | 31 | 1.69 | Not shortlisted |
| SYSTRAN PNMT API | Adjusted NMT | 33 | 0.774 | 34 | 1.83 | Not shortlisted |
| ModernMT Enterprise Edition | Adjusted NMT | 34 | 0.770 | 32 | 1.75 | Not shortlisted |
| Baidu Translate API | Off-the-shelf NMT | 35 | 0.756 | 35 | 2.24 | Not shortlisted |
| Google Cloud Advanced Translation API | Adjusted NMT | 36 | 0.754 | 38 | 2.37 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 37 | 0.753 | 37 | 2.25 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 38 | 0.746 | 36 | 2.24 | Not shortlisted |
| Naver Papago NMT Commercial | Off-the-shelf NMT | 39 | 0.725 | 39 | 2.45 | Not shortlisted |
| Cohere Tiny Aya | Adjusted LLM | 40 | 0.719 | 40 | 2.82 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 41 | 0.592 | 41 | 5.57 | Not shortlisted |
| TartuNLP | Off-the-shelf NMT | 42 | 0.319 | 42 | 14.04 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.809 | — | 1.08 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.812 | — | 1.12 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.803 | — | 1.28 | 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 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.
Some solutions are on the list through one metric alone, with a low rank on the other. Google Gemini 3.5 Flash is there through MetricX, at 12th on COMET-Kiwi, and OpenAI GPT-5.5 through COMET-Kiwi, at 17th on MetricX. The one-step solutions on both meet general translation quality, and the one on DeepL Next Gen, first on both metrics, misses it, with four major errors. The model nearest the cut is Anthropic Claude Sonnet 4.6, 11th on COMET-Kiwi and 12th on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | Google Gemini 3.5 Flash, one-step LLM adjusted for Standard requirements | $3.05 |
| Standard | Google Gemini 3.5 Flash, two-step | $3.05 |
| Advanced | Google Gemini 3.1 Pro Preview, two-step with the gender-neutrality agent | $28.91 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.13 | Open-weight LLMs |
| Mistral Large 3 | $0.82 | LLMs |
| Mistral Medium 3.1 | $1 | LLMs |
| OpenAI GPT-4.1 | $2.86 | LLMs |
| Google Gemini 3.5 Flash | $3.05 | LLMs |
| Yandex Cloud Translate v2 | $3.94 | MT engines |
| Google Gemini 3.1 Pro Preview | $4.14 | LLMs |
| Anthropic Claude Sonnet 5 | $5.38 | LLMs |
| OpenAI GPT-5.6 Sol | $6.98 | LLMs |
| OpenAI GPT-5.5 | $10.1 | LLMs |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $27.06 | LLMs |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Model | Tokens per character | Class |
|---|---|---|
| AI21 Jamba Large 1.7 | 0.543 | LLMs |
| Anthropic Claude Fable 5 | 0.417 | LLMs |
| Anthropic Claude Opus 4.8 | 0.416 | LLMs |
| Anthropic Claude Sonnet 5 | 0.416 | LLMs |
| Mistral Medium 3.1 | 0.401 | LLMs |
| Mistral Large 3 | 0.396 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.386 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.386 | LLMs |
| MoonshotAI Kimi K2.5 | 0.362 | LLMs |
| Widn AI Tower Sugarloaf | 0.324 | LLM-based MT engines |
| Alibaba Qwen3-235B-A22B | 0.323 | Open-weight LLMs |
| MiniMax M2.5 | 0.298 | LLMs |
| DeepSeek V3.2 | 0.297 | LLMs |
| Cohere Tiny Aya | 0.273 | LLMs |
| z.ai GLM-5 | 0.272 | LLMs |
| Cohere Command A Translate | 0.271 | LLMs |
| OpenAI GPT-5.5 | 0.268 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.267 | LLMs |
| Google Gemini 3.5 Flash | 0.267 | LLMs |
| Google Gemma 4 | 0.267 | Open-weight LLMs |
| OpenAI GPT-4.1 | 0.267 | LLMs |
| OpenAI GPT-5.6 Terra | 0.266 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.265 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.265 | LLMs |
| OpenAI GPT-5.6 Sol | 0.265 | LLMs |
| OpenAI GPT-5.6 Luna | 0.264 | LLMs |
The tokenizers produce a median 0.29 tokens per character of Russian, 1.2 times their count for English, and the Anthropic tokenizers produce the most on the shortlist, 0.42 for Anthropic Claude Fable 5 and Anthropic Claude Sonnet 5. The post-edit step raises the cost of the solution on Anthropic Claude Fable 5 from $27.06 to $59.73 per million characters. On Google Gemini 3.5 Flash the cost stays at $3.05 per million characters, so meeting the Standard requirements costs no more than meeting the Basic ones.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | OpenAI GPT-5.6 Sol, one-step LLM adjusted for Standard requirements | 44 |
| Standard | OpenAI GPT-5.6 Sol, one-step LLM adjusted for Standard requirements | 44 |
| Advanced | Google Gemini 3.1 Pro Preview, two-step with the gender-neutrality agent | 201 |
| 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 |
| Pangeanic PangeaMT | 0.6 | MT engines |
| DeepL Next Gen | 1.1 | LLM-based MT engines |
| SYSTRAN PNMT API | 1.1 | MT engines |
| ModernMT Enterprise Edition | 1.2 | MT engines |
| Translated Lara 2 | 1.3 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 2.3 | Open-weight LLMs |
| Amazon Translate | 3 | MT engines |
| Yandex Cloud Translate v2 | 3.9 | MT engines |
| TartuNLP | 5.6 | MT engines |
| Tilde TildeOpen LLM engine | 8.1 | LLM-based MT engines |
| OpenAI GPT-5.6 Luna | 12 | LLMs |
| Baidu Translate API | 13 | MT engines |
| Cohere Tiny Aya | 14 | LLMs |
| Supertext LLM | 15 | LLM-based MT engines |
| Alibaba Machine Translation API | 19 | MT engines |
| OpenAI GPT-5.6 Terra | 19 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 19 | LLMs |
| OpenAI GPT-4.1 | 20 | LLMs |
| Anthropic Claude Haiku 4.5 | 28 | LLMs |
| Mistral Medium 3.1 | 31 | LLMs |
| Cohere Command A Translate | 32 | LLMs |
| MiniMax M2.5 | 36 | LLMs |
| Anthropic Claude Sonnet 5 | 40 | LLMs |
| Mistral Large 3 | 42 | LLMs |
| Anthropic Claude Fable 5 | 44 | LLMs |
| OpenAI GPT-5.6 Sol | 44 | LLMs |
| OpenAI GPT-5.5 | 46 | LLMs |
| Anthropic Claude Opus 4.8 | 50 | LLMs |
| Google Gemini 3.5 Flash | 54 | LLMs |
| Anthropic Claude Sonnet 4.6 | 58 | LLMs |
| NiuTrans | 58 | MT engines |
| Widn AI Tower Sugarloaf | 59 | LLM-based MT engines |
| DeepSeek V3.2 | 61 | LLMs |
| xAI Grok 4.20 (reasoning) | 66 | LLMs |
| AI21 Jamba Large 1.7 | 75 | LLMs |
| Naver Papago NMT Commercial | 77 | MT engines |
| Alibaba Qwen3-235B-A22B | 92 | Open-weight LLMs |
| AppTek NMT | 103 | MT engines |
| Google Gemini 3.1 Pro Preview | 106 | LLMs |
| MoonshotAI Kimi K2.5 | 137 | LLMs |
| z.ai GLM-5 | 151 | LLMs |
| Google Gemma 4 | 166 | Open-weight 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 Russian.
Nine of the 10 real-time models in Russian are MT engines or LLM-based MT engines, and the other one is the open-weight OpenAI gpt-oss-120b, which did not make the shortlist. The two real-time solutions on the shortlist, on DeepL Next Gen and Translated Lara 2, have four and 12 major errors, more than general translation quality allows. The fastest solution that meets every Basic and Standard requirement takes 44 seconds per 1,000 words, and the only one that meets every Advanced requirement takes 201.
| 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 | 99.1% | ✗ | $63.32 | 96 |
| Google Gemini 3.1 Pro Preview | two-step, with the gender-neutrality agent | 100% | ✓ | $28.91 | 201 |
For gender neutrality in Russian we check neutral rephrasing, plurals and impersonal forms. None of the one-step solutions meet it, with a median 2.5 gender-neutrality errors per 1,000 words. With the gender-neutrality agent, the weighted errors of the two-step solution on Anthropic Claude Fable 5 fall from 12.9 to 11.2 per 1,000 words, and those of the one on Google Gemini 3.1 Pro Preview rise from 15.8 to 18.1.
The agreement of each pair of automatic metrics is above its threshold in Russian: 0.59 for the similarity pair against 0.46, and 0.66 for the quality-estimation pair against 0.53. The three LQA judges agree on a median of 96% of the sentences of the shortlisted solutions, so the results on this page rest on measurement we can rely on.
Leave feedback