The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into German 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 (Sie) 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 German, with a column for each requirement you select. Under Optimize for, choose which of them comes first: the one with the fewest total errors, the lowest cost or the shortest time. Click a column heading to sort by it.
The shortlisted solutions on 11 models, with their Intento LQA results. Total errors per 1,000 words add the weighted general translation quality errors (minor 1, major 5, critical 10) to the missed glossary terms, sentences in the wrong register, tag errors and sentences that fail gender neutrality, each counted once. Formatting counts as met or missed and does not add to the total errors. The models that did not make the shortlist have automatic scores only: the shortlist. 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 in German, with 17.7 to 20.5 total errors per 1,000 words. The one-step solution on Google Gemini 3.1 Pro Preview costs $4.07 per million characters and takes 79 seconds per 1,000 words. The two-step solutions (translation + post-edit) on OpenAI GPT-5.6 Luna and DeepSeek V3.2 cost $0.91 and $9.57 per million characters and take 30 and 166 seconds per 1,000 words. Optimize for selects Google Gemini 3.1 Pro Preview for quality and OpenAI GPT-5.6 Luna for cost and latency.
Under the Standard requirements, a limit to one step leaves only the solution on Google Gemini 3.1 Pro Preview. The three solutions cost $0.91 per million characters or more and take 30 seconds per 1,000 words or more, so a lower cost cap or time limit removes them all. None of the three real-time (<2 sec) solutions on the shortlist, on DeepL Next Gen, Microsoft Translator and Amazon Translate, meet tag handling or formatting. Under the Basic requirements, real time leaves the solution on DeepL Next Gen.
The two-step solution (translation + post-edit) on OpenAI GPT-5.6 Luna, the cheapest under the Standard requirements at $0.91 per million characters, has no result with the gender-neutrality agent, so it drops out under the Advanced ones. With the agent, the two-step solution on DeepSeek V3.2 costs $9.34 per million characters against $9.57 and takes 159 seconds per 1,000 words against 166. Its total errors fall from 19.9 to 15.8 per 1,000 words, and its weighted errors rise from 11.9 to 15.8 per 1,000 words.
Anthropic Claude Sonnet 5
three-step
OpenAI GPT-5.6 Luna
two-step
DeepL Next Gen
one-step NMT adjusted for Standard requirements
Google Gemini 3.1 Pro Preview
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Luna
two-step
OpenAI GPT-5.6 Luna
two-step
Google Gemini 3.1 Pro Preview
one-step with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
one-step with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
one-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 3.1 Pro Preview | 17.7 | $4.07 | 79 |
| Standard | Tier 1: simple solution + backup | ★ ☆☆ | one-step | Google Gemini 3.1 Pro Preview | 17.7 | $4.07 | 79 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | with the gender-neutrality agent | Google Gemini 3.1 Pro Preview | 12.5 | $4.02 | 94 |
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 four of the seven solutions that meet the Basic ones. Tag handling and formatting remove the one-step solutions on Anthropic Claude Opus 4.8 and DeepL Next Gen. Terminology removes the three-step solution (translation + LQA + repair) on Anthropic Claude Sonnet 5, and terminology and tag handling the one on OpenAI GPT-4.1. Under the Basic requirements, Optimize for selects the one on Anthropic Claude Sonnet 5 for quality, with 16.6 total errors per 1,000 words at $579 per million characters.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| OpenAI GPT-5.6 Luna | 0.4 | 18.8 | Misses a Standard requirement |
| DeepSeek V3.2 | 0.86 | 20.1 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 2.7 | 30 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.07 | 10.5 | Meets every Standard requirement |
| Anthropic Claude Sonnet 5 | 6.28 | 22.1 | Misses a Standard requirement |
| Microsoft Translator | 10 | 71.2 | Misses a Standard requirement |
| Amazon Translate | 15 | 76.1 | Misses a Standard requirement |
| Anthropic Claude Opus 4.8 | 15.85 | 11.9 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 15.8 | Misses a Standard requirement |
Among the one-step solutions in German, the one on Google Gemini 3.1 Pro Preview has the fewest errors, 10.5 weighted errors per 1,000 words, at $4.07 per million characters. None of the five that cost more has fewer errors. The ones on Anthropic Claude Opus 4.8 and DeepL Next Gen cost $15.85 and $27.50 per million characters and have 11.9 and 15.8 weighted errors per 1,000 words. The ones on Microsoft Translator and Amazon Translate cost $10 and $15 per million characters and have 71.2 and 76.1 weighted errors per 1,000 words.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 22.1 |
| Gender neutrality | 8.1 |
| Tag handling | 1.3 |
| Tone of voice | 0.2 |
| Terminology | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
The post-edit step fixes two near misses. The one-step solution on OpenAI GPT-5.6 Luna has four major errors, just over what general translation quality allows, and 97.6% of the tags right. Its two-step version (translation + post-edit) has two major errors and every tag right. The other, on DeepSeek V3.2, has one critical error, four major errors, 20.1 weighted errors per 1,000 words and 98.8% of the terms right. Its two-step version has no critical error, two major errors, 11.9 weighted errors per 1,000 words and every term right.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in German | Class |
|---|---|---|
| Amazon Translate | 76.1 | MT engines |
| Microsoft Translator | 71.2 | MT engines |
| Meta Llama 4 Maverick | 36.9 | Open-weight LLMs |
| OpenAI GPT-4.1 | 30 | LLMs |
| Cohere Command A Translate | 28.7 | LLMs |
| Anthropic Claude Sonnet 5 | 22.1 | LLMs |
| DeepSeek V3.2 | 20.1 | LLMs |
| OpenAI GPT-5.6 Luna | 18.8 | LLMs |
| DeepL Next Gen | 15.8 | LLM-based MT engines |
| Anthropic Claude Opus 4.8 | 11.9 | LLMs |
| Google Gemini 3.1 Pro Preview | 10.5 | LLMs |
No translation into German: Widn AI Tower Vesuvius (cut off part-way).
The one-step solutions on LLMs, open-weight included, have 10.5 to 36.9 weighted errors per 1,000 words in German, the one on the LLM-based MT engine DeepL Next Gen 15.8, and the ones on the MT engines Microsoft Translator and Amazon Translate 71.2 and 76.1. Translations from DeepL Next Gen have 1.5 times as many weighted errors as translations from Google Gemini 3.1 Pro Preview. All eight shortlisted LLMs support German by our measured criteria, and the solutions on two of them meet general translation quality.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Google Gemma 4 | 0.768 | 10 | 0.87 | 21 | No | Not shortlisted |
| Widn AI Tower Sugarloaf | 0.768 | 11 | 0.84 | 14 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.768 | 12 | 0.83 | 13 | No | Not shortlisted |
| Supertext LLM | 0.766 | 14 | 0.83 | 12 | No | Not shortlisted |
| Tilde TildeOpen LLM engine | 0.766 | 15 | 0.88 | 24 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.765 | 16 | 0.88 | 23 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.765 | 17 | 0.91 | 28 | No | Not shortlisted |
| OpenAI GPT-5.5 | 0.764 | 18 | 0.82 | 11 | No | Not shortlisted |
| Anthropic Claude Fable 5 | 0.764 | 19 | 0.99 | 34 | No | Not shortlisted |
| z.ai GLM-5 | 0.764 | 20 | 0.86 | 18 | No | Not shortlisted |
| Google Gemini 3.5 Flash | 0.763 | 22 | 0.97 | 33 | No | Not shortlisted |
| Mistral Large 3 | 0.762 | 23 | 0.85 | 15 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.762 | 24 | 0.95 | 31 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.762 | 25 | 0.96 | 32 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.762 | 26 | 0.86 | 20 | No | Not shortlisted |
| Translated Lara 2 | 0.761 | 27 | 0.90 | 27 | No | Not shortlisted |
| Google Translation LLM | 0.761 | 28 | 1.26 | 41 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.759 | 29 | 0.86 | 17 | No | Not shortlisted |
| MiniMax M2.5 | 0.759 | 30 | 0.95 | 30 | No | Not shortlisted |
| Pangeanic PangeaMT | 0.755 | 31 | 1.03 | 36 | No | Not shortlisted |
| iFlyTek MT | 0.755 | 32 | 0.88 | 22 | No | Not shortlisted |
| Cohere Tiny Aya | 0.754 | 33 | 0.89 | 25 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.754 | 34 | 0.93 | 29 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.753 | 35 | 0.86 | 19 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.751 | 36 | 0.90 | 26 | No | Not shortlisted |
| SpeakLeash Bielik-11B-v3.0-Instruct | 0.743 | 37 | 1.22 | 40 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.739 | 38 | 1.04 | 37 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.736 | 39 | 1.02 | 35 | No | Not shortlisted |
| Baidu Translate API | 0.733 | 40 | 1.09 | 39 | No | Not shortlisted |
| Oracle Cloud Language Service | 0.727 | 41 | 1.07 | 38 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.719 | 42 | 1.85 | 44 | No | Not shortlisted |
| Tencent Cloud TMT | 0.710 | 43 | 1.42 | 43 | No | Not shortlisted |
| Naver Papago NMT Commercial | 0.694 | 44 | 1.42 | 42 | No | Not shortlisted |
| AppTek NMT | 0.675 | 45 | 2.21 | 45 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.666 | 46 | 2.44 | 46 | No | Not shortlisted |
| Widn AI Tower Anthill | 0.429 | 47 | 10.44 | 48 | No | Not shortlisted |
| TartuNLP | 0.390 | 48 | 10.11 | 47 | No | Not shortlisted |
| DeepL Next Gen | 0.779 | 1 | 0.73 | 2 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Luna | 0.773 | 2 | 0.79 | 5 | Yes | Top 8 by both metrics |
| Cohere Command A Translate | 0.772 | 3 | 0.80 | 6 | Yes | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | 0.772 | 4 | 0.82 | 10 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | 0.772 | 5 | 0.78 | 4 | Yes | Top 8 by both metrics |
| OpenAI GPT-4.1 | 0.770 | 6 | 0.81 | 9 | Yes | Top 8 by COMET-Kiwi |
| DeepSeek V3.2 | 0.769 | 7 | 0.85 | 16 | Yes | Top 8 by COMET-Kiwi |
| Microsoft Translator | 0.769 | 8 | 0.73 | 1 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.768 | 9 | 0.80 | 8 | Yes | Top 8 by MetricX |
| Amazon Translate | 0.766 | 13 | 0.74 | 3 | Yes | Top 8 by MetricX |
| Meta Llama 4 Maverick | 0.764 | 21 | 0.80 | 7 | Yes | Top 8 by MetricX |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| DeepL Next Gen | Adjusted NMT | 1 | 0.779 | 2 | 0.73 | Top 8 by both metrics |
| OpenAI GPT-5.6 Luna | Adjusted LLM | 2 | 0.773 | 5 | 0.79 | Top 8 by both metrics |
| Cohere Command A Translate | Adjusted LLM | 3 | 0.772 | 6 | 0.80 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 4 | 0.772 | 10 | 0.82 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 5 | 0.772 | 4 | 0.78 | Top 8 by both metrics |
| OpenAI GPT-4.1 | Adjusted LLM | 6 | 0.770 | 9 | 0.81 | Top 8 by COMET-Kiwi |
| DeepSeek V3.2 | Adjusted LLM | 7 | 0.769 | 16 | 0.85 | Top 8 by COMET-Kiwi |
| Microsoft Translator | Off-the-shelf NMT | 8 | 0.769 | 1 | 0.73 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 9 | 0.768 | 8 | 0.80 | Top 8 by MetricX |
| Google Gemma 4 | Adjusted LLM | 10 | 0.768 | 21 | 0.87 | Not shortlisted |
| Widn AI Tower Sugarloaf | Adjusted NMT | 11 | 0.768 | 14 | 0.84 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 12 | 0.768 | 13 | 0.83 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 13 | 0.766 | 3 | 0.74 | Top 8 by MetricX |
| Supertext LLM | Off-the-shelf NMT | 14 | 0.766 | 12 | 0.83 | Not shortlisted |
| Tilde TildeOpen LLM engine | Adjusted NMT | 15 | 0.766 | 24 | 0.88 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 16 | 0.765 | 23 | 0.88 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Off-the-shelf LLM | 17 | 0.765 | 28 | 0.91 | Not shortlisted |
| OpenAI GPT-5.5 | Adjusted LLM | 18 | 0.764 | 11 | 0.82 | Not shortlisted |
| Anthropic Claude Fable 5 | Adjusted LLM | 19 | 0.764 | 34 | 0.99 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 20 | 0.764 | 18 | 0.86 | Not shortlisted |
| Meta Llama 4 Maverick | Adjusted LLM | 21 | 0.764 | 7 | 0.80 | Top 8 by MetricX |
| Google Gemini 3.5 Flash | Adjusted LLM | 22 | 0.763 | 33 | 0.97 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 23 | 0.762 | 15 | 0.85 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 24 | 0.762 | 31 | 0.95 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 25 | 0.762 | 32 | 0.96 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 26 | 0.762 | 20 | 0.86 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 27 | 0.761 | 27 | 0.90 | Not shortlisted |
| Google Translation LLM | Off-the-shelf NMT | 28 | 0.761 | 41 | 1.26 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 29 | 0.759 | 17 | 0.86 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 30 | 0.759 | 30 | 0.95 | Not shortlisted |
| Pangeanic PangeaMT | Off-the-shelf NMT | 31 | 0.755 | 36 | 1.03 | Not shortlisted |
| iFlyTek MT | Off-the-shelf NMT | 32 | 0.755 | 22 | 0.88 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 33 | 0.754 | 25 | 0.89 | Not shortlisted |
| ModernMT Enterprise Edition | Adjusted NMT | 34 | 0.754 | 29 | 0.93 | Not shortlisted |
| SYSTRAN PNMT API | Adjusted NMT | 35 | 0.753 | 19 | 0.86 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 36 | 0.751 | 26 | 0.90 | Not shortlisted |
| SpeakLeash Bielik-11B-v3.0-Instruct | Adjusted LLM | 37 | 0.743 | 40 | 1.22 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 38 | 0.739 | 37 | 1.04 | Not shortlisted |
| Yandex Cloud Translate v2 | Adjusted NMT | 39 | 0.736 | 35 | 1.02 | Not shortlisted |
| Baidu Translate API | Off-the-shelf NMT | 40 | 0.733 | 39 | 1.09 | Not shortlisted |
| Oracle Cloud Language Service | Off-the-shelf NMT | 41 | 0.727 | 38 | 1.07 | Not shortlisted |
| Google Cloud Advanced Translation API | Adjusted NMT | 42 | 0.719 | 44 | 1.85 | Not shortlisted |
| Tencent Cloud TMT | Off-the-shelf NMT | 43 | 0.710 | 43 | 1.42 | Not shortlisted |
| Naver Papago NMT Commercial | Off-the-shelf NMT | 44 | 0.694 | 42 | 1.42 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 45 | 0.675 | 45 | 2.21 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 46 | 0.666 | 46 | 2.44 | Not shortlisted |
| Widn AI Tower Anthill | Adjusted NMT | 47 | 0.429 | 48 | 10.44 | Not shortlisted |
| TartuNLP | Off-the-shelf NMT | 48 | 0.390 | 47 | 10.11 | Not shortlisted |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.772 | — | 0.77 | Same model family as OpenAI GPT-5.6 Luna |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.773 | — | 0.78 | Same model family as OpenAI GPT-5.6 Luna |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.761 | — | 1.02 | 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 48 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.
Of the three one-step solutions that meet general translation quality, the one on Anthropic Claude Opus 4.8 is on the shortlist through COMET-Kiwi alone, 10th on MetricX, and the one on Google Gemini 3.1 Pro Preview through MetricX alone. The third, on DeepL Next Gen, is in the top 8 of both. Amazon Translate is on the list through MetricX alone too, 3rd there, and its translations have 76.1 weighted errors per 1,000 words. The models nearest the cut are Google Gemma 4, 10th on COMET-Kiwi, and OpenAI GPT-5.5, 11th on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.6 Luna, two-step | $0.91 |
| Standard | OpenAI GPT-5.6 Luna, two-step | $0.91 |
| Advanced | Google Gemini 3.1 Pro Preview, one-step with the gender-neutrality agent | $4.02 |
| Model | US dollars per million characters | Class |
|---|---|---|
| OpenAI GPT-5.6 Luna | $0.4 | LLMs |
| DeepSeek V3.2 | $0.86 | LLMs |
| OpenAI GPT-4.1 | $2.7 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.07 | LLMs |
| Anthropic Claude Sonnet 5 | $6.28 | LLMs |
| Microsoft Translator | $10 | MT engines |
| Amazon Translate | $15 | MT engines |
| Anthropic Claude Opus 4.8 | $15.85 | LLMs |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Opus 4.8 | 0.481 | LLMs |
| Anthropic Claude Fable 5 | 0.48 | LLMs |
| Anthropic Claude Sonnet 5 | 0.478 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.35 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.347 | LLMs |
| Mistral Medium 3.1 | 0.346 | LLMs |
| Mistral Large 3 | 0.338 | LLMs |
| AI21 Jamba Large 1.7 | 0.311 | LLMs |
| MoonshotAI Kimi K2.5 | 0.308 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.288 | Open-weight LLMs |
| Widn AI Tower Sugarloaf | 0.286 | LLM-based MT engines |
| DeepSeek V3.2 | 0.285 | LLMs |
| Widn AI Tower Anthill | 0.283 | LLM-based MT engines |
| xAI Grok 4.20 (reasoning) | 0.272 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.271 | LLMs |
| Cohere Command A Translate | 0.267 | LLMs |
| z.ai GLM-5 | 0.265 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.253 | LLMs |
| Google Gemini 3.5 Flash | 0.252 | LLMs |
| Google Gemma 4 | 0.252 | Open-weight LLMs |
| MiniMax M2.5 | 0.252 | LLMs |
| Cohere Tiny Aya | 0.245 | LLMs |
| OpenAI GPT-4.1 | 0.242 | LLMs |
| OpenAI GPT-5.5 | 0.241 | LLMs |
| OpenAI GPT-5.6 Terra | 0.241 | LLMs |
| OpenAI GPT-5.6 Sol | 0.24 | LLMs |
| OpenAI GPT-5.6 Luna | 0.239 | LLMs |
The tokenizers produce a median 0.27 tokens per character of German, 1.2 times their count for English. The two-step solution (translation + post-edit) on DeepSeek V3.2 costs $9.57 per million characters, against $0.86 in one step. The three-step solutions (translation + LQA + repair) on Anthropic Claude Sonnet 5 and OpenAI GPT-4.1 cost $579 and $593 per million characters, against $6.28 and $2.70 in one step.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | DeepL Next Gen, one-step NMT adjusted for Standard requirements | 0.7 |
| Standard | OpenAI GPT-5.6 Luna, two-step | 30 |
| Advanced | Google Gemini 3.1 Pro Preview, one-step with the gender-neutrality agent | 94 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.2 | MT engines |
| Google Cloud Advanced Translation API | 0.2 | MT engines |
| Pangeanic PangeaMT | 0.5 | MT engines |
| Google Translation LLM | 0.6 | LLM-based MT engines |
| DeepL Next Gen | 0.7 | LLM-based MT engines |
| Translated Lara 2 | 1.1 | LLM-based MT engines |
| Amazon Translate | 1.3 | MT engines |
| OpenAI gpt-oss-120b | 1.3 | Open-weight LLMs |
| ModernMT Enterprise Edition | 3.1 | MT engines |
| Yandex Cloud Translate v2 | 3.4 | MT engines |
| Oracle Cloud Language Service | 3.8 | MT engines |
| SYSTRAN PNMT API | 6.9 | MT engines |
| TartuNLP | 6.9 | MT engines |
| Tilde TildeOpen LLM engine | 7 | LLM-based MT engines |
| Widn AI Tower Anthill | 9 | LLM-based MT engines |
| OpenAI GPT-5.6 Luna | 12 | LLMs |
| Google Gemini 3.5 Flash | 13 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 13 | LLMs |
| Baidu Translate API | 14 | MT engines |
| Supertext LLM | 15 | LLM-based MT engines |
| OpenAI GPT-4.1 | 16 | LLMs |
| Cohere Tiny Aya | 19 | LLMs |
| Alibaba Machine Translation API | 21 | MT engines |
| OpenAI GPT-5.6 Terra | 21 | LLMs |
| Anthropic Claude Haiku 4.5 | 26 | LLMs |
| Cohere Command A Translate | 30 | LLMs |
| Mistral Medium 3.1 | 32 | LLMs |
| Meta Llama 4 Maverick | 33 | Open-weight LLMs |
| MiniMax M2.5 | 35 | LLMs |
| Swiss AI Apertus-8B-2509 | 36 | Open-weight LLMs |
| Anthropic Claude Sonnet 5 | 39 | LLMs |
| Mistral Large 3 | 39 | LLMs |
| DeepSeek V3.2 | 40 | LLMs |
| OpenAI GPT-5.5 | 44 | LLMs |
| AI21 Jamba Large 1.7 | 46 | LLMs |
| Widn AI Tower Sugarloaf | 47 | LLM-based MT engines |
| Anthropic Claude Fable 5 | 47 | LLMs |
| Anthropic Claude Opus 4.8 | 49 | LLMs |
| iFlyTek MT | 52 | MT engines |
| Anthropic Claude Sonnet 4.6 | 53 | LLMs |
| SpeakLeash Bielik-11B-v3.0-Instruct | 54 | Open-weight LLMs |
| Tencent Cloud TMT | 56 | MT engines |
| OpenAI GPT-5.6 Sol | 56 | LLMs |
| xAI Grok 4.20 (reasoning) | 64 | LLMs |
| Naver Papago NMT Commercial | 67 | MT engines |
| Google Gemini 3.1 Pro Preview | 79 | LLMs |
| AppTek NMT | 83 | MT engines |
| Alibaba Qwen3-235B-A22B | 115 | Open-weight LLMs |
| Google Gemma 4 | 177 | Open-weight LLMs |
| z.ai GLM-5 | 181 | LLMs |
| MoonshotAI Kimi K2.5 | 290 | 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 German.
Nine of the ten real-time (<2 sec) models in German are MT engines or LLM-based MT engines, and the other is the open-weight OpenAI gpt-oss-120b, which is not on the shortlist. The fastest solution takes 0.7 seconds per 1,000 words under the Basic requirements, on DeepL Next Gen, 30 under the Standard ones, the two-step solution (translation + post-edit) on OpenAI GPT-5.6 Luna, and 94 under the Advanced ones, on Google Gemini 3.1 Pro Preview with the gender-neutrality agent.
| 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 | one-step, with the gender-neutrality agent | 100% | ✓ | $4.02 | 94 |
| DeepSeek V3.2 | two-step, with the gender-neutrality agent | 100% | ✓ | $9.34 | 159 |
For tone of voice in German we check the formal register (Sie). Translations from the one-step solutions have 100% of the sentences in it, except Microsoft Translator with 91.5% and Amazon Translate with 95.7%. For gender neutrality we check neutral rephrasing, paired forms used sparingly, no gender glyphs and no default gender. The one-step solutions have a median 8.1 gender-neutrality errors per 1,000 words. With the gender-neutrality agent, the two best solutions have 100% of the sentences about people gender-neutral.
In German, the agreement of both pairs of automatic metrics is above the threshold, and neither pair is borderline. The three LQA judges agree on a median of 94.4% of the sentences of the shortlisted solutions, so the Intento LQA results of these solutions rest on measurement we can rely on.
Leave feedback