The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Estonian 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 (teie (teie-vorm)) 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 Estonian, 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.
None of the three solutions that meet every Standard requirement is ahead on errors, cost and time at once. The one-step solution on OpenAI GPT-5.6 Sol has 19.4 total errors per 1,000 words, costs $7.53 per million characters and takes 79 seconds per 1,000 words. The three-step solution (translation + LQA + repair) on Anthropic Claude Fable 5 has 10.9 total errors per 1,000 words, at $641 per million characters and 238 seconds per 1,000 words. Optimize for selects the first for cost and latency and the second for quality.
For the Standard requirements, a cost cap from $7.54 to under $641 per million characters, a time limit from 79 to 237 seconds per 1,000 words or a limit of one or two steps leaves only the one-step solution on OpenAI GPT-5.6 Sol. With a cost cap of at most $7.53 or a time limit of at most 78 seconds, no solution meets them. A cap from 12.5 to under 19.4 total errors per 1,000 words works the other way. It leaves only the three-step solutions (translation + LQA + repair) on Anthropic Claude Fable 5 and Google Gemini 3.1 Pro Preview.
Both solutions that meet the Advanced requirements do so without the gender-neutrality agent, with 100% of the sentences about people gender-neutral, so the agent does not add to the cost or the time. They are the one-step solution on OpenAI GPT-5.6 Sol and the three-step solution on Anthropic Claude Fable 5, and Optimize for selects them as under the Standard requirements: the first for cost and latency, the second for quality.
Anthropic Claude Fable 5
three-step
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
three-step
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
three-step
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
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 Sol | 19.4 | $7.53 | 79 |
| Standard | Tier 1: simple solution + backup | ★ ☆☆ | one-step | OpenAI GPT-5.6 Sol | 19.4 | $7.53 | 79 |
| Advanced | Tier 1: simple solution + backup | ★ ☆ | one-step | OpenAI GPT-5.6 Sol | 19.4 | $7.53 | 79 |
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 two of the five solutions that meet the Basic ones, the three-step solutions (translation + LQA + repair) on Anthropic Claude Opus 4.8 and xAI Grok 4.20 (reasoning). They have 92.6% and 92.0% of the glossary terms right, and the one on Anthropic Claude Opus 4.8 has 99.6% of the tags right. Only the one-step solution on OpenAI GPT-5.6 Sol and the three-step solution on Anthropic Claude Fable 5 meet the Advanced requirements. Optimize for selects the same solutions at all three levels.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| xAI Grok 4.20 (reasoning) | 1.2 | 50.8 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 3.06 | 44.8 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.54 | 36.6 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.77 | 27 | Misses a Standard requirement |
| Anthropic Claude Sonnet 5 | 5.92 | 55.7 | Misses a Standard requirement |
| Anthropic Claude Sonnet 4.6 | 6.98 | 32.6 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 7.53 | 19.4 | Meets every Standard requirement |
| Microsoft Translator | 10 | 87.7 | Misses a Standard requirement |
| Anthropic Claude Opus 4.8 | 14.85 | 33.9 | Misses a Standard requirement |
| Google Translation LLM | 20.04 | 49.8 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 49.8 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 29.62 | 21.8 | Misses a Standard requirement |
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 40.7 |
| Tag handling | 1 |
| Tone of voice | 0.8 |
| Terminology | 0.5 |
| Gender neutrality | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
General translation quality is the requirement most one-step solutions miss, and every three-step solution (translation + LQA + repair) meets it. On Google Gemini 3.1 Pro Preview, the weighted errors fall from 27.0 per 1,000 words in one step to 20.8 in the two-step solution (translation + post-edit) and to 12.5 in three steps. On Anthropic Claude Fable 5, they fall from 21.8 per 1,000 words in one step to 10.9 in three steps. The one-step LLM solutions that miss terminology have 94.4% to 98.7% of the glossary terms right.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Estonian | Class |
|---|---|---|
| Microsoft Translator | 87.7 | MT engines |
| Anthropic Claude Sonnet 5 | 55.7 | LLMs |
| xAI Grok 4.20 (reasoning) | 50.8 | LLMs |
| DeepL Next Gen | 49.8 | LLM-based MT engines |
| Google Translation LLM | 49.8 | LLM-based MT engines |
| OpenAI GPT-4.1 | 44.8 | LLMs |
| Google Gemini 3.5 Flash | 36.6 | LLMs |
| Anthropic Claude Opus 4.8 | 33.9 | LLMs |
| Anthropic Claude Sonnet 4.6 | 32.6 | LLMs |
| Google Gemini 3.1 Pro Preview | 27 | LLMs |
| Anthropic Claude Fable 5 | 21.8 | LLMs |
| OpenAI GPT-5.6 Sol | 19.4 | LLMs |
No translation into Estonian: Widn AI Tower Anthill (far too short, wrong language), Widn AI Tower Vesuvius (far too short, wrong language).
The one-step solutions on the nine LLMs have 19.4 to 55.7 weighted errors per 1,000 words, those on the two LLM-based MT engines, DeepL Next Gen and Google Translation LLM, 49.8 each, and the one on the MT engine Microsoft Translator 87.7. Only the translations from Google Translation LLM and Microsoft Translator have critical errors, 3 and 1. All nine shortlisted LLMs meet our language-support criteria in Estonian. The text we got back from Widn AI Tower Anthill and Widn AI Tower Vesuvius is far too short and in the wrong language.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| z.ai GLM-5 | 0.860 | 9 | 2.27 | 12 | No | Not shortlisted |
| DeepSeek V3.2 | 0.857 | 11 | 2.16 | 10 | No | Not shortlisted |
| Tilde TildeOpen LLM engine | 0.849 | 14 | 2.38 | 18 | No | Not shortlisted |
| Translated Lara 2 | 0.848 | 16 | 2.34 | 17 | No | Not shortlisted |
| OpenAI GPT-5.5 | 0.847 | 17 | 2.33 | 16 | No | Not shortlisted |
| Mistral Large 3 | 0.844 | 18 | 2.29 | 14 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.843 | 19 | 2.60 | 23 | No | Not shortlisted |
| SpeakLeash Bielik-11B-v3.0-Instruct | 0.843 | 20 | 2.39 | 19 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.842 | 21 | 2.31 | 15 | No | Not shortlisted |
| Pangeanic PangeaMT | 0.841 | 22 | 2.68 | 24 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.833 | 23 | 2.43 | 20 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.833 | 24 | 2.48 | 21 | No | Not shortlisted |
| MiniMax M2.5 | 0.830 | 25 | 2.93 | 27 | No | Not shortlisted |
| Google Gemma 4 | 0.827 | 26 | 3.04 | 29 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.821 | 27 | 2.76 | 25 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.819 | 28 | 3.15 | 30 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.815 | 29 | 3.00 | 28 | No | Not shortlisted |
| Cohere Command A Translate | 0.809 | 30 | 2.77 | 26 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.806 | 31 | 3.18 | 31 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.801 | 32 | 3.60 | 33 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.798 | 33 | 3.26 | 32 | No | Not shortlisted |
| Amazon Translate | 0.791 | 34 | 3.98 | 35 | No | Not shortlisted |
| Cohere Tiny Aya | 0.774 | 35 | 3.79 | 34 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.758 | 36 | 4.12 | 36 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.711 | 37 | 5.71 | 37 | No | Not shortlisted |
| TartuNLP | 0.552 | 38 | 9.79 | 38 | No | Not shortlisted |
| AppTek NMT | 0.285 | 39 | 17.32 | 39 | No | Not shortlisted |
| Google Gemini 3.1 Pro Preview | 0.870 | 1 | 1.86 | 2 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | 0.866 | 2 | 1.90 | 3 | Yes | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | 0.865 | 3 | 2.20 | 11 | Yes | Top 8 by COMET-Kiwi |
| DeepL Next Gen | 0.864 | 4 | 1.78 | 1 | Yes | Top 8 by both metrics |
| xAI Grok 4.20 (reasoning) | 0.862 | 5 | 2.10 | 8 | Yes | Top 8 by both metrics |
| Google Gemini 3.5 Flash | 0.861 | 6 | 2.10 | 7 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 4.6 | 0.861 | 7 | 2.12 | 9 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | 0.861 | 8 | 2.08 | 5 | Yes | Top 8 by both metrics |
| OpenAI GPT-4.1 | 0.860 | 10 | 1.99 | 4 | Yes | Top 8 by MetricX |
| Anthropic Claude Fable 5 | 0.857 | 12 | 2.09 | 6 | Yes | Top 8 by MetricX |
| Google Translation LLM | 0.853 | 13 | 2.28 | 13 | Yes | Real-time place |
| Microsoft Translator | 0.848 | 15 | 2.49 | 22 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 1 | 0.870 | 2 | 1.86 | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | Adjusted LLM | 2 | 0.866 | 3 | 1.90 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 3 | 0.865 | 11 | 2.20 | Top 8 by COMET-Kiwi |
| DeepL Next Gen | Adjusted NMT | 4 | 0.864 | 1 | 1.78 | Top 8 by both metrics |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 5 | 0.862 | 8 | 2.10 | Top 8 by both metrics |
| Google Gemini 3.5 Flash | Adjusted LLM | 6 | 0.861 | 7 | 2.10 | Top 8 by both metrics |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 7 | 0.861 | 9 | 2.12 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 8 | 0.861 | 5 | 2.08 | Top 8 by both metrics |
| z.ai GLM-5 | Adjusted LLM | 9 | 0.860 | 12 | 2.27 | Not shortlisted |
| OpenAI GPT-4.1 | Adjusted LLM | 10 | 0.860 | 4 | 1.99 | Top 8 by MetricX |
| DeepSeek V3.2 | Adjusted LLM | 11 | 0.857 | 10 | 2.16 | Not shortlisted |
| Anthropic Claude Fable 5 | Adjusted LLM | 12 | 0.857 | 6 | 2.09 | Top 8 by MetricX |
| Google Translation LLM | Off-the-shelf NMT | 13 | 0.853 | 13 | 2.28 | Real-time place |
| Tilde TildeOpen LLM engine | Adjusted NMT | 14 | 0.849 | 18 | 2.38 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 15 | 0.848 | 22 | 2.49 | Real-time place |
| Translated Lara 2 | Adjusted NMT | 16 | 0.848 | 17 | 2.34 | Not shortlisted |
| OpenAI GPT-5.5 | Adjusted LLM | 17 | 0.847 | 16 | 2.33 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 18 | 0.844 | 14 | 2.29 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 19 | 0.843 | 23 | 2.60 | Not shortlisted |
| SpeakLeash Bielik-11B-v3.0-Instruct | Adjusted LLM | 20 | 0.843 | 19 | 2.39 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 21 | 0.842 | 15 | 2.31 | Not shortlisted |
| Pangeanic PangeaMT | Off-the-shelf NMT | 22 | 0.841 | 24 | 2.68 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 23 | 0.833 | 20 | 2.43 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 24 | 0.833 | 21 | 2.48 | Not shortlisted |
| MiniMax M2.5 | Off-the-shelf LLM | 25 | 0.830 | 27 | 2.93 | Not shortlisted |
| Google Gemma 4 | Adjusted LLM | 26 | 0.827 | 29 | 3.04 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 27 | 0.821 | 25 | 2.76 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 28 | 0.819 | 30 | 3.15 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 29 | 0.815 | 28 | 3.00 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 30 | 0.809 | 26 | 2.77 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 31 | 0.806 | 31 | 3.18 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 32 | 0.801 | 33 | 3.60 | Not shortlisted |
| SYSTRAN PNMT API | Adjusted NMT | 33 | 0.798 | 32 | 3.26 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 34 | 0.791 | 35 | 3.98 | Not shortlisted |
| Cohere Tiny Aya | Adjusted LLM | 35 | 0.774 | 34 | 3.79 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 36 | 0.758 | 36 | 4.12 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 37 | 0.711 | 37 | 5.71 | Not shortlisted |
| TartuNLP | Off-the-shelf NMT | 38 | 0.552 | 38 | 9.79 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 39 | 0.285 | 39 | 17.32 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.857 | — | 2.04 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.866 | — | 2.06 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.858 | — | 2.10 | 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 39 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. Translations from DeepL Next Gen, first on MetricX, have 49.8 weighted errors per 1,000 words by Intento LQA, and those from Google Gemini 3.1 Pro Preview, first on COMET-Kiwi, 27.0. Anthropic Claude Fable 5 made the shortlist through MetricX alone, 12th on COMET-Kiwi, and its one-step solution has 21.8 weighted errors per 1,000 words, second only to OpenAI GPT-5.6 Sol. The models nearest the cut are z.ai GLM-5, 9th on COMET-Kiwi, and DeepSeek V3.2, 10th on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.6 Sol, one-step LLM adjusted for Standard requirements | $7.53 |
| Standard | OpenAI GPT-5.6 Sol, one-step LLM adjusted for Standard requirements | $7.53 |
| Advanced | OpenAI GPT-5.6 Sol, one-step LLM adjusted for Standard requirements | $7.53 |
| Model | US dollars per million characters | Class |
|---|---|---|
| xAI Grok 4.20 (reasoning) | $1.2 | LLMs |
| OpenAI GPT-4.1 | $3.06 | LLMs |
| Google Gemini 3.5 Flash | $3.54 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.77 | LLMs |
| Anthropic Claude Sonnet 5 | $5.92 | LLMs |
| Anthropic Claude Sonnet 4.6 | $6.98 | LLMs |
| OpenAI GPT-5.6 Sol | $7.53 | LLMs |
| Microsoft Translator | $10 | MT engines |
| Anthropic Claude Opus 4.8 | $14.85 | LLMs |
| Google Translation LLM | $20.04 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $29.62 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Opus 4.8 | 0.504 | LLMs |
| Anthropic Claude Sonnet 5 | 0.502 | LLMs |
| Anthropic Claude Fable 5 | 0.501 | LLMs |
| Mistral Medium 3.1 | 0.431 | LLMs |
| Mistral Large 3 | 0.427 | LLMs |
| AI21 Jamba Large 1.7 | 0.409 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.398 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.396 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.381 | Open-weight LLMs |
| DeepSeek V3.2 | 0.378 | LLMs |
| Cohere Command A Translate | 0.375 | LLMs |
| z.ai GLM-5 | 0.371 | LLMs |
| MoonshotAI Kimi K2.5 | 0.364 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.334 | LLMs |
| Google Gemini 3.5 Flash | 0.334 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.334 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.334 | LLMs |
| Google Gemma 4 | 0.329 | Open-weight LLMs |
| MiniMax M2.5 | 0.32 | LLMs |
| OpenAI GPT-5.6 Terra | 0.313 | LLMs |
| OpenAI GPT-4.1 | 0.312 | LLMs |
| OpenAI GPT-5.6 Luna | 0.312 | LLMs |
| OpenAI GPT-5.6 Sol | 0.312 | LLMs |
| OpenAI GPT-5.5 | 0.311 | LLMs |
| Cohere Tiny Aya | 0.295 | LLMs |
| 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 | 79 |
| Standard | OpenAI GPT-5.6 Sol, one-step LLM adjusted for Standard requirements | 79 |
| Advanced | OpenAI GPT-5.6 Sol, one-step LLM adjusted for Standard requirements | 79 |
| 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 | 0.5 | LLM-based MT engines |
| Pangeanic PangeaMT | 0.5 | MT engines |
| DeepL Next Gen | 0.6 | LLM-based MT engines |
| ModernMT Enterprise Edition | 0.9 | MT engines |
| Amazon Translate | 1.1 | MT engines |
| SYSTRAN PNMT API | 1.2 | MT engines |
| Translated Lara 2 | 1.5 | LLM-based MT engines |
| Yandex Cloud Translate v2 | 2.8 | MT engines |
| OpenAI gpt-oss-120b | 3.7 | Open-weight LLMs |
| TartuNLP | 5.4 | MT engines |
| Tilde TildeOpen LLM engine | 7 | LLM-based MT engines |
| OpenAI GPT-5.6 Luna | 15 | LLMs |
| Google Gemini 3.5 Flash | 17 | LLMs |
| Cohere Tiny Aya | 17 | LLMs |
| OpenAI GPT-4.1 | 18 | LLMs |
| Alibaba Machine Translation API | 25 | MT engines |
| OpenAI GPT-5.6 Terra | 31 | LLMs |
| DeepSeek V3.2 | 32 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 34 | LLMs |
| Mistral Medium 3.1 | 43 | LLMs |
| Anthropic Claude Haiku 4.5 | 43 | LLMs |
| Swiss AI Apertus-8B-2509 | 51 | Open-weight LLMs |
| AI21 Jamba Large 1.7 | 51 | LLMs |
| MiniMax M2.5 | 55 | LLMs |
| Cohere Command A Translate | 58 | LLMs |
| OpenAI GPT-5.5 | 63 | LLMs |
| SpeakLeash Bielik-11B-v3.0-Instruct | 63 | Open-weight LLMs |
| Mistral Large 3 | 67 | LLMs |
| Anthropic Claude Sonnet 5 | 69 | LLMs |
| AppTek NMT | 75 | MT engines |
| OpenAI GPT-5.6 Sol | 79 | LLMs |
| Anthropic Claude Fable 5 | 79 | LLMs |
| Anthropic Claude Sonnet 4.6 | 82 | LLMs |
| Anthropic Claude Opus 4.8 | 95 | LLMs |
| xAI Grok 4.20 (reasoning) | 96 | LLMs |
| Google Gemini 3.1 Pro Preview | 112 | LLMs |
| Alibaba Qwen3-235B-A22B | 138 | Open-weight LLMs |
| Google Gemma 4 | 195 | Open-weight LLMs |
| z.ai GLM-5 | 263 | 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 Estonian.
The fastest one-step LLM solutions on the shortlist, on Google Gemini 3.5 Flash and OpenAI GPT-4.1, take 17 and 18 seconds per 1,000 words, against 79 for the one on OpenAI GPT-5.6 Sol. Both miss general translation quality, with 36.6 and 44.8 weighted errors per 1,000 words against 19.4.
| 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 |
|---|---|---|---|---|---|
| OpenAI GPT-5.6 Sol | one-step | 100% | ✓ | $7.53 | 79 |
| Anthropic Claude Fable 5 | three-step | 100% | ✓ | $641 | 238 |
For tone of voice we check the formal register, with teie. Translations from all nine shortlisted LLMs have 100% of the sentences in the formal register, against 93.6% for DeepL Next Gen, 87.2% for Google Translation LLM and 59.6% for Microsoft Translator. Estonian has no grammatical gender for persons, so for gender neutrality we check that a translation avoids gendered terms and does not add a gender the source leaves open.
Neither pair of automatic metrics is borderline in Estonian: each pair agrees above its threshold even at the low end of its range. The three LQA judges agree on a median of 95.5% of the sentences across the 12 shortlisted solutions. So neither the agreement of the metrics nor that of the LQA judges limits the results on this page.
Leave feedback