The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Hebrew 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 Hebrew. |
| Tone of voice | The formal register 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 Hebrew, 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. 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.
Three solutions meet every Standard requirement, and none leads on errors, cost and time at once. The three-step solution (translation + LQA + repair) on Google Gemini 3.1 Pro Preview has the fewest total errors, 26.6 per 1,000 words, at $540 per million characters and 287 seconds per 1,000 words. The one-step solution on Anthropic Claude Fable 5 has 39.1 total errors per 1,000 words at $30.72 per million characters and 86 seconds per 1,000 words. Optimizing for quality selects the first, and for cost or latency the second.
Under the Standard requirements, a cost cap of at most $515 per million characters or a time limit of at most 286 seconds per 1,000 words leaves the one-step solution on Anthropic Claude Fable 5 alone, and so does a limit of one or two steps. None of the three fits a cap of at most $30.71 per million characters or a limit of at most 85 seconds per 1,000 words. The three-step solution (translation + LQA + repair) on z.ai GLM-5 costs $516 per million characters and takes the longest, 461 seconds per 1,000 words.
For the Advanced requirements, the gender-neutrality agent works with the three-step solutions (translation + LQA + repair) on Google Gemini 3.1 Pro Preview and z.ai GLM-5. On Google Gemini 3.1 Pro Preview, total errors fall from 26.6 to 20.9 per 1,000 words and the time goes from 287 to 300 seconds per 1,000 words. On z.ai GLM-5, total errors rise from 40.1 to 41.5 per 1,000 words and the cost goes from $516 to $915 per million characters.
Google Gemini 3.1 Pro Preview
three-step
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Google Gemini 3.1 Pro Preview
three-step
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Google Gemini 3.1 Pro Preview
three-step with the gender-neutrality agent
Misses 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 | Anthropic Claude Fable 5 | 39.1 | $30.72 | 86 |
| Standard | Tier 1: simple solution + backup | ★ ☆☆ | one-step | Anthropic Claude Fable 5 | 39.1 | $30.72 | 86 |
| 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.
The one-step solution on Anthropic Claude Fable 5 and the three-step solutions (translation + LQA + repair) on Google Gemini 3.1 Pro Preview and z.ai GLM-5 meet every Basic and Standard requirement. All three meet tone of voice, tag handling and formatting, so the Standard requirements keep them, and Optimize for selects the same ones at both levels. The Advanced requirements remove all three. Closest is the three-step solution on Google Gemini 3.1 Pro Preview with the gender-neutrality agent, at 20.9 total errors per 1,000 words.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Mistral Large 3 | 1.19 | 44.5 | Misses a Standard requirement |
| xAI Grok 4.20 (reasoning) | 1.27 | 27.7 | Misses a Standard requirement |
| z.ai GLM-5 | 3.26 | 34.9 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.98 | 28.3 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 7.27 | 20.4 | Misses a Standard requirement |
| Anthropic Claude Sonnet 4.6 | 8.5 | 37.6 | Misses a Standard requirement |
| Microsoft Translator | 10 | 63.6 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 33.6 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 30.72 | 16.5 | Meets every Standard requirement |
The cost and errors front has three LLM solutions, the one-step solution on Anthropic Claude Fable 5 and the three-step solutions (translation + LQA + repair) on z.ai GLM-5 and Google Gemini 3.1 Pro Preview. Of the one-step solutions with a cost and a conclusive Intento LQA result, the one on Anthropic Claude Fable 5 costs the most, $30.72 per million characters, and has the fewest weighted errors, 16.5 per 1,000 words. The next dearest, on DeepL Next Gen, costs $27.50 per million characters, with 33.6 weighted errors per 1,000 words.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 34.9 |
| Gender neutrality | 22.6 |
| Tone of voice | 0 |
| Tag handling | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
Of tone of voice, tag handling, formatting and general translation quality, the fewest one-step solutions in Hebrew meet general translation quality. None of the two-step solutions (translation + post-edit) meets it. The two-step solution on OpenAI GPT-5.6 Sol has 19.1 weighted errors per 1,000 words but five major errors. With three steps (translation + LQA + repair), weighted errors fall from 28.3 to 6.6 per 1,000 words on Google Gemini 3.1 Pro Preview and from 34.9 to 13.8 on z.ai GLM-5.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Hebrew | Class |
|---|---|---|
| OpenAI gpt-oss-120b | 75.8 | Open-weight LLMs |
| Microsoft Translator | 63.6 | MT engines |
| Cohere Command A Translate | 45.5 | LLMs |
| Mistral Large 3 | 44.5 | LLMs |
| Anthropic Claude Sonnet 4.6 | 37.6 | LLMs |
| z.ai GLM-5 | 34.9 | LLMs |
| DeepL Next Gen | 33.6 | LLM-based MT engines |
| Google Gemini 3.1 Pro Preview | 28.3 | LLMs |
| xAI Grok 4.20 (reasoning) | 27.7 | LLMs |
| OpenAI GPT-5.6 Sol | 20.4 | LLMs |
| Anthropic Claude Fable 5 | 16.5 | LLMs |
No translation into Hebrew: Widn AI Tower Anthill (far too short, wrong language, wrong script), Widn AI Tower Vesuvius (far too short, wrong language, wrong script).
In one step, translations from the eight shortlisted LLMs that are not open-weight and have a conclusive Intento LQA result have 16.5 to 45.5 weighted errors per 1,000 words in Hebrew. Translations from the LLM-based MT engine DeepL Next Gen have 33.6 weighted errors per 1,000 words, and those from the MT engine Microsoft Translator 63.6, with 19 major errors. The output from Widn AI Tower Anthill and Widn AI Tower Vesuvius is not a translation into Hebrew: it is far too short, in the wrong language and in the wrong script.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Google Gemini 3.5 Flash | 0.796 | 11 | 1.14 | 12 | No | Not shortlisted |
| Anthropic Claude Sonnet 5 | 0.795 | 12 | 1.15 | 13 | No | Not shortlisted |
| DeepSeek V3.2 | 0.794 | 13 | 1.13 | 10 | No | Not shortlisted |
| Anthropic Claude Opus 4.8 | 0.794 | 15 | 1.13 | 11 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.793 | 17 | 1.25 | 19 | No | Not shortlisted |
| OpenAI GPT-5.5 | 0.793 | 18 | 1.21 | 17 | No | Not shortlisted |
| Translated Lara 2 | 0.792 | 19 | 1.34 | 23 | No | Not shortlisted |
| Google Gemma 4 | 0.792 | 20 | 1.26 | 20 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.790 | 21 | 1.16 | 15 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.788 | 22 | 1.15 | 14 | No | Not shortlisted |
| MiniMax M2.5 | 0.786 | 23 | 1.30 | 22 | No | Not shortlisted |
| Google Translation LLM | 0.784 | 24 | 1.68 | 27 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.781 | 25 | 1.50 | 25 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.776 | 26 | 1.37 | 24 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.766 | 27 | 1.62 | 26 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.764 | 28 | 2.20 | 31 | No | Not shortlisted |
| Amazon Translate | 0.763 | 29 | 1.70 | 28 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.743 | 30 | 1.84 | 29 | No | Not shortlisted |
| Oracle Cloud Language Service | 0.741 | 31 | 1.99 | 30 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.741 | 32 | 2.57 | 33 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.740 | 33 | 2.25 | 32 | No | Not shortlisted |
| Cohere Tiny Aya | 0.690 | 34 | 3.50 | 37 | No | Not shortlisted |
| NiuTrans | 0.663 | 35 | 2.84 | 34 | No | Not shortlisted |
| AppTek NMT | 0.660 | 36 | 3.42 | 35 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.650 | 37 | 3.43 | 36 | No | Not shortlisted |
| DeepL Next Gen | 0.804 | 1 | 1.08 | 4 | Yes | Top 8 by both metrics |
| Cohere Command A Translate | 0.802 | 2 | 1.09 | 6 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | 0.800 | 3 | 1.02 | 1 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 4.6 | 0.800 | 4 | 1.07 | 3 | Yes | Top 8 by both metrics |
| OpenAI GPT-4.1 | 0.800 | 5 | 1.10 | 7 | Yes | Top 8 by both metrics |
| xAI Grok 4.20 (reasoning) | 0.798 | 6 | 1.28 | 21 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Fable 5 | 0.798 | 7 | 1.13 | 9 | Yes | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | 0.796 | 8 | 1.08 | 5 | Yes | Top 8 by both metrics |
| Mistral Large 3 | 0.796 | 9 | 1.05 | 2 | Yes | Top 8 by MetricX |
| z.ai GLM-5 | 0.796 | 10 | 1.11 | 8 | Yes | Top 8 by MetricX |
| Microsoft Translator | 0.794 | 14 | 1.21 | 18 | Yes | Real-time place |
| OpenAI gpt-oss-120b | 0.793 | 16 | 1.20 | 16 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| DeepL Next Gen | Off-the-shelf NMT | 1 | 0.804 | 4 | 1.08 | Top 8 by both metrics |
| Cohere Command A Translate | Adjusted LLM | 2 | 0.802 | 6 | 1.09 | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | Adjusted LLM | 3 | 0.800 | 1 | 1.02 | Top 8 by both metrics |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 4 | 0.800 | 3 | 1.07 | Top 8 by both metrics |
| OpenAI GPT-4.1 | Adjusted LLM | 5 | 0.800 | 7 | 1.10 | Top 8 by both metrics |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 6 | 0.798 | 21 | 1.28 | Top 8 by COMET-Kiwi |
| Anthropic Claude Fable 5 | Adjusted LLM | 7 | 0.798 | 9 | 1.13 | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 8 | 0.796 | 5 | 1.08 | Top 8 by both metrics |
| Mistral Large 3 | Adjusted LLM | 9 | 0.796 | 2 | 1.05 | Top 8 by MetricX |
| z.ai GLM-5 | Adjusted LLM | 10 | 0.796 | 8 | 1.11 | Top 8 by MetricX |
| Google Gemini 3.5 Flash | Adjusted LLM | 11 | 0.796 | 12 | 1.14 | Not shortlisted |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 12 | 0.795 | 13 | 1.15 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 13 | 0.794 | 10 | 1.13 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 14 | 0.794 | 18 | 1.21 | Real-time place |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 15 | 0.794 | 11 | 1.13 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 16 | 0.793 | 16 | 1.20 | Real-time place |
| Mistral Medium 3.1 | Adjusted LLM | 17 | 0.793 | 19 | 1.25 | Not shortlisted |
| OpenAI GPT-5.5 | Adjusted LLM | 18 | 0.793 | 17 | 1.21 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 19 | 0.792 | 23 | 1.34 | Not shortlisted |
| Google Gemma 4 | Adjusted LLM | 20 | 0.792 | 20 | 1.26 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 21 | 0.790 | 15 | 1.16 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 22 | 0.788 | 14 | 1.15 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 23 | 0.786 | 22 | 1.30 | Not shortlisted |
| Google Translation LLM | Adjusted NMT | 24 | 0.784 | 27 | 1.68 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 25 | 0.781 | 25 | 1.50 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 26 | 0.776 | 24 | 1.37 | Not shortlisted |
| SYSTRAN PNMT API | Off-the-shelf NMT | 27 | 0.766 | 26 | 1.62 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 28 | 0.764 | 31 | 2.20 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 29 | 0.763 | 28 | 1.70 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 30 | 0.743 | 29 | 1.84 | Not shortlisted |
| Oracle Cloud Language Service | Off-the-shelf NMT | 31 | 0.741 | 30 | 1.99 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Off-the-shelf LLM | 32 | 0.741 | 33 | 2.57 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 33 | 0.740 | 32 | 2.25 | Not shortlisted |
| Cohere Tiny Aya | Adjusted LLM | 34 | 0.690 | 37 | 3.50 | Not shortlisted |
| NiuTrans | Off-the-shelf NMT | 35 | 0.663 | 34 | 2.84 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 36 | 0.660 | 35 | 3.42 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 37 | 0.650 | 36 | 3.43 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.796 | — | 1.10 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.798 | — | 1.14 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.795 | — | 1.12 | 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 37 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 shortlisted solutions in Hebrew differently. Translations from DeepL Next Gen, first on COMET-Kiwi, have 33.6 weighted errors per 1,000 words in one step. Mistral Large 3 is on the shortlist through MetricX alone, second there and ninth on COMET-Kiwi, and translations from it have 44.5 weighted errors per 1,000 words. Nearest the cut in the tail are Google Gemini 3.5 Flash, 11th 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 | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | $30.72 |
| Standard | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | $30.72 |
| Advanced | No solution | — |
| Model | US dollars per million characters | Class |
|---|---|---|
| Mistral Large 3 | $1.19 | LLMs |
| xAI Grok 4.20 (reasoning) | $1.27 | LLMs |
| OpenAI GPT-4.1 | $2.94 | LLMs |
| z.ai GLM-5 | $3.26 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.98 | LLMs |
| OpenAI GPT-5.6 Sol | $7.27 | LLMs |
| Anthropic Claude Sonnet 4.6 | $8.5 | LLMs |
| Microsoft Translator | $10 | MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $30.72 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Mistral Medium 3.1 | 0.827 | LLMs |
| Mistral Large 3 | 0.822 | LLMs |
| z.ai GLM-5 | 0.769 | LLMs |
| Anthropic Claude Sonnet 5 | 0.632 | LLMs |
| Anthropic Claude Opus 4.8 | 0.631 | LLMs |
| Anthropic Claude Fable 5 | 0.63 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.597 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.596 | LLMs |
| MoonshotAI Kimi K2.5 | 0.521 | LLMs |
| MiniMax M2.5 | 0.465 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.455 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.451 | LLMs |
| AI21 Jamba Large 1.7 | 0.432 | LLMs |
| Google Gemini 3.5 Flash | 0.426 | LLMs |
| DeepSeek V3.2 | 0.424 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.424 | LLMs |
| Google Gemma 4 | 0.422 | Open-weight LLMs |
| Alibaba Qwen3-235B-A22B | 0.379 | Open-weight LLMs |
| Cohere Command A Translate | 0.377 | LLMs |
| OpenAI GPT-4.1 | 0.366 | LLMs |
| OpenAI GPT-5.6 Terra | 0.363 | LLMs |
| OpenAI GPT-5.5 | 0.362 | LLMs |
| OpenAI GPT-5.6 Sol | 0.362 | LLMs |
| OpenAI GPT-5.6 Luna | 0.361 | LLMs |
| Cohere Tiny Aya | 0.36 | LLMs |
With billing on tokens, the tokenizers produce a median 0.43 tokens per character of the Hebrew translation, 1.8 times their count for English. The tokenizer of Anthropic Claude Fable 5 produces 0.63 tokens per character, and that of OpenAI GPT-5.6 Sol 0.36. On Google Gemini 3.1 Pro Preview, one step costs $4.98 per million characters, two steps (translation + post-edit) $30.33 and three steps (translation + LQA + repair) $540.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | 86 |
| Standard | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | 86 |
| Advanced | No solution | — |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.4 | MT engines |
| ModernMT Enterprise Edition | 0.4 | MT engines |
| Google Cloud Advanced Translation API | 0.5 | MT engines |
| DeepL Next Gen | 0.9 | LLM-based MT engines |
| Google Translation LLM | 1.3 | LLM-based MT engines |
| Translated Lara 2 | 1.5 | LLM-based MT engines |
| Amazon Translate | 1.5 | MT engines |
| SYSTRAN PNMT API | 1.8 | MT engines |
| OpenAI gpt-oss-120b | 2.4 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 2.8 | MT engines |
| Oracle Cloud Language Service | 6.1 | MT engines |
| OpenAI GPT-5.6 Luna | 14 | LLMs |
| Cohere Tiny Aya | 15 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 17 | LLMs |
| Alibaba Machine Translation API | 18 | MT engines |
| OpenAI GPT-4.1 | 31 | LLMs |
| Cohere Command A Translate | 33 | LLMs |
| Mistral Medium 3.1 | 40 | LLMs |
| Swiss AI Apertus-8B-2509 | 44 | Open-weight LLMs |
| OpenAI GPT-5.6 Terra | 44 | LLMs |
| xAI Grok 4.20 (reasoning) | 45 | LLMs |
| DeepSeek V3.2 | 49 | LLMs |
| MiniMax M2.5 | 52 | LLMs |
| Anthropic Claude Haiku 4.5 | 54 | LLMs |
| NiuTrans | 59 | MT engines |
| Mistral Large 3 | 60 | LLMs |
| Google Gemini 3.5 Flash | 63 | LLMs |
| OpenAI GPT-5.5 | 69 | LLMs |
| OpenAI GPT-5.6 Sol | 82 | LLMs |
| Google Gemini 3.1 Pro Preview | 84 | LLMs |
| Anthropic Claude Sonnet 5 | 85 | LLMs |
| Anthropic Claude Fable 5 | 86 | LLMs |
| AI21 Jamba Large 1.7 | 87 | LLMs |
| Anthropic Claude Sonnet 4.6 | 100 | LLMs |
| AppTek NMT | 104 | MT engines |
| Anthropic Claude Opus 4.8 | 108 | LLMs |
| Alibaba Qwen3-235B-A22B | 131 | Open-weight LLMs |
| Google Gemma 4 | 169 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 182 | LLMs |
| z.ai GLM-5 | 253 | 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 Hebrew.
The nine real-time models in Hebrew are five MT engines, three LLM-based MT engines and the open-weight OpenAI gpt-oss-120b. Translations from the three of them on the shortlist, DeepL Next Gen, Microsoft Translator and OpenAI gpt-oss-120b, have 33.6, 63.6 and 75.8 weighted errors per 1,000 words, over the bar of general translation quality. The solutions that meet every Standard requirement take 86 to 461 seconds per 1,000 words.
| The two best solutions | Solution | Sentences about people that are gender-neutral | Meets every Advanced requirement | Cost per million characters | Seconds per 1,000 words |
|---|---|---|---|---|---|
| Google Gemini 3.1 Pro Preview | three-step, with the gender-neutrality agent | 97.2% | ✗ | $919 | 300 |
| z.ai GLM-5 | three-step, with the gender-neutrality agent | 68.8% | ✗ | $915 | 457 |
For Hebrew we check the formal register and gender-neutral wording by neutral rephrasing, with no default gender. With the gender-neutrality agent, the three-step solution (translation + LQA + repair) on z.ai GLM-5 has 29.7 weighted errors per 1,000 words, over the bar for general translation quality. Without the agent, the same solution has 13.8 weighted errors per 1,000 words. The one on Google Gemini 3.1 Pro Preview has 19.8 weighted errors per 1,000 words with the agent, just under the bar.
The three LQA judges agree on a median of 94.7% of the sentences of the 12 shortlisted solutions in Hebrew, and both pairs of automatic metrics agree above the threshold. So the Intento LQA error counts on this page rest on measurement we can rely on. Only the one-step solution on OpenAI GPT-4.1 has no conclusive Intento LQA result.
Leave feedback