The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Javanese 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 Javanese. |
| Tone of voice | The formal register (krama) 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. |
| Solution | Sentences on which the three judges agree (%) | Status |
|---|---|---|
| AI21 Jamba Large 1.7 | 92.1% | Passes (90% or more) |
| DeepL Next Gen | 90.4% | Passes (90% or more) |
| Anthropic Claude Haiku 4.5 | 89.3% | Does not pass |
| Google Gemini 3.1 Pro Preview | 86.4% | Does not pass |
| Google Gemma 4 | 86.4% | Does not pass |
| Cohere Command A Translate | 86.4% | Does not pass |
| OpenAI GPT-5.5 | 85.9% | Does not pass |
| Google Gemini 3.5 Flash | 85.3% | Does not pass |
| Swiss AI Apertus-8B-2509 | 85.3% | Does not pass |
| z.ai GLM-5 | 83.6% | Does not pass |
| OpenAI GPT-4.1 | 81.9% | Does not pass |
| Anthropic Claude Sonnet 4.6 | 80.2% | Does not pass |
| xAI Grok 4.20 (non-reasoning) | 76.8% | Does not pass |
In Javanese the agreement of the similarity pair, 0.545 to 0.571, and of the quality-estimation pair, 0.553 to 0.581, stays above the minimum of each, 0.461 and 0.534, so the automatic scores can rank the models within Javanese. They are not a requirement result. The two solutions that pass the agreement check are on AI21 Jamba Large 1.7, at 92.1% of the sentences, and DeepL Next Gen, at 90.4%. Of the rest, the three judges agree most often on the translations from Anthropic Claude Haiku 4.5, on 89.3% of the sentences, short of 90%.
Pick a quality level, or your own requirements, bars and constraints, to see the shortlisted solutions that meet them in Javanese, with a column for each requirement you select. Under Optimize for, choose which of them comes first: the one with the lowest cost or the shortest time. Click a column heading to sort by it.
The shortlisted solutions on 13 models, with their results on tone of voice, tags and formatting, which do not depend on the agreement of the three LQA judges, and on general translation quality only where the judges agree. We cannot assess translation quality in Javanese: why. 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.
None of the shortlisted solutions in Javanese meet more than two of tone of voice, tag handling and formatting. The one on Anthropic Claude Sonnet 4.6 meets tone of voice and tag handling, at $7.35 per million characters and 80 seconds per 1,000 words. The three that meet tag handling and formatting, on z.ai GLM-5, Anthropic Claude Haiku 4.5 and Cohere Command A Translate, have 2.1% to 74.5% of the sentences we check in the formal register.
Only two shortlisted solutions in Javanese cost less than $1.84 per million characters, on Google Gemma 4 at $0.14 and xAI Grok 4.20 (non-reasoning) at $1.18, and of tone of voice, tag handling and formatting both meet formatting alone. A time limit of 30 seconds per 1,000 words leaves three, on DeepL Next Gen at 0.8 seconds and on xAI Grok 4.20 (non-reasoning) and OpenAI GPT-4.1 at 20 and 23. None of these three meet tag handling or tone of voice, and the one on DeepL Next Gen also misses formatting.
The shortlisted solutions on AI21 Jamba Large 1.7 and OpenAI GPT-5.5 come close on tone of voice and tag handling: their translations have 97.9% of the sentences we check in the formal register and keep 99.2% and 99.6% of the tags. The solution on AI21 Jamba Large 1.7 meets formatting, at $3.72 per million characters and 51 seconds per 1,000 words.
We cannot assess general translation quality in Javanese: how confident we are.
We cannot assess general translation quality in Javanese: how confident we are.
We cannot assess general translation quality in Javanese: how confident we are.
| Model | COMET-Kiwi scores of the shortlisted solutions in Javanese (higher is better) | Class |
|---|---|---|
| AI21 Jamba Large 1.7 | 0.791 | LLMs |
| Cohere Command A Translate | 0.76 | LLMs |
| Swiss AI (Apertus) Swiss AI Apertus-8B-2509 | 0.758 | Open-weight LLMs |
| Anthropic Claude Haiku 4.5 | 0.741 | LLMs |
| DeepL Next Gen | 0.711 | LLM-based MT engines |
| Google Gemini 3.5 Flash | 0.708 | LLMs |
| z.ai GLM-5 | 0.708 | LLMs |
| OpenAI GPT-4.1 | 0.706 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.705 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.703 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.702 | LLMs |
| OpenAI GPT-5.5 | 0.689 | LLMs |
| Google Gemma 4 | 0.684 | Open-weight LLMs |
The automatic scores rank models within Javanese and are not a requirement result. The top 8 on COMET-Kiwi are seven LLMs, one of them the open-weight Swiss AI Apertus-8B-2509, and one LLM-based MT engine, DeepL Next Gen, which is 2nd on MetricX. The other LLM-based MT engine, Translated Lara 2, is 10th on COMET-Kiwi and 17th on MetricX. The four MT engines we got translations from are 18th to 30th on COMET-Kiwi, led by Google Cloud Advanced Translation API.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Translated Lara 2 | 0.705 | 10 | 1.97 | 17 | No | Not shortlisted |
| OpenAI GPT-5.6 Luna | 0.695 | 13 | 1.80 | 9 | No | Not shortlisted |
| Anthropic Claude Fable 5 | 0.678 | 16 | 1.91 | 14 | No | Not shortlisted |
| Mistral Large 3 | 0.676 | 17 | 1.87 | 13 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.674 | 18 | 2.67 | 24 | No | Not shortlisted |
| Anthropic Claude Sonnet 5 | 0.674 | 19 | 2.11 | 19 | No | Not shortlisted |
| Anthropic Claude Opus 4.8 | 0.674 | 20 | 2.12 | 20 | No | Not shortlisted |
| DeepSeek V3.2 | 0.671 | 21 | 2.05 | 18 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.669 | 22 | 3.01 | 25 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.665 | 23 | 2.44 | 22 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.656 | 24 | 2.43 | 21 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.655 | 25 | 2.55 | 23 | No | Not shortlisted |
| MiniMax M2.5 | 0.641 | 26 | 4.39 | 28 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.638 | 27 | 3.05 | 26 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.633 | 28 | 3.13 | 27 | No | Not shortlisted |
| Cohere Tiny Aya | 0.437 | 29 | 10.36 | 29 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.331 | 30 | 11.10 | 30 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.791 | 1 | 1.35 | 1 | Yes | Top 8 by both metrics |
| Cohere Command A Translate | 0.760 | 2 | 1.83 | 11 | Yes | Top 8 by COMET-Kiwi |
| Swiss AI Apertus-8B-2509 | 0.758 | 3 | 1.74 | 4 | Yes | Top 8 by both metrics |
| Anthropic Claude Haiku 4.5 | 0.741 | 4 | 1.94 | 16 | Yes | Top 8 by COMET-Kiwi |
| DeepL Next Gen | 0.711 | 5 | 1.66 | 2 | Yes | Top 8 by both metrics |
| Google Gemini 3.5 Flash | 0.708 | 6 | 1.82 | 10 | Yes | Top 8 by COMET-Kiwi |
| z.ai GLM-5 | 0.708 | 7 | 1.94 | 15 | Yes | Top 8 by COMET-Kiwi |
| OpenAI GPT-4.1 | 0.706 | 8 | 1.86 | 12 | Yes | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | 0.705 | 9 | 1.73 | 3 | Yes | Top 8 by MetricX |
| Anthropic Claude Sonnet 4.6 | 0.703 | 11 | 1.76 | 5 | Yes | Top 8 by MetricX |
| xAI Grok 4.20 (non-reasoning) | 0.702 | 12 | 1.77 | 7 | Yes | Top 8 by MetricX |
| OpenAI GPT-5.5 | 0.689 | 14 | 1.77 | 6 | Yes | Top 8 by MetricX |
| Google Gemma 4 | 0.684 | 15 | 1.80 | 8 | Yes | Top 8 by MetricX |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| AI21 Jamba Large 1.7 | Adjusted LLM | 1 | 0.791 | 1 | 1.35 | Top 8 by both metrics |
| Cohere Command A Translate | Adjusted LLM | 2 | 0.760 | 11 | 1.83 | Top 8 by COMET-Kiwi |
| Swiss AI Apertus-8B-2509 | Off-the-shelf LLM | 3 | 0.758 | 4 | 1.74 | Top 8 by both metrics |
| Anthropic Claude Haiku 4.5 | Off-the-shelf LLM | 4 | 0.741 | 16 | 1.94 | Top 8 by COMET-Kiwi |
| DeepL Next Gen | Off-the-shelf NMT | 5 | 0.711 | 2 | 1.66 | Top 8 by both metrics |
| Google Gemini 3.5 Flash | Off-the-shelf LLM | 6 | 0.708 | 10 | 1.82 | Top 8 by COMET-Kiwi |
| z.ai GLM-5 | Off-the-shelf LLM | 7 | 0.708 | 15 | 1.94 | Top 8 by COMET-Kiwi |
| OpenAI GPT-4.1 | Adjusted LLM | 8 | 0.706 | 12 | 1.86 | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | Off-the-shelf LLM | 9 | 0.705 | 3 | 1.73 | Top 8 by MetricX |
| Translated Lara 2 | Adjusted NMT | 10 | 0.705 | 17 | 1.97 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 11 | 0.703 | 5 | 1.76 | Top 8 by MetricX |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | 12 | 0.702 | 7 | 1.77 | Top 8 by MetricX |
| OpenAI GPT-5.6 Luna | Adjusted LLM | 13 | 0.695 | 9 | 1.80 | Not shortlisted |
| OpenAI GPT-5.5 | Adjusted LLM | 14 | 0.689 | 6 | 1.77 | Top 8 by MetricX |
| Google Gemma 4 | Adjusted LLM | 15 | 0.684 | 8 | 1.80 | Top 8 by MetricX |
| Anthropic Claude Fable 5 | Adjusted LLM | 16 | 0.678 | 14 | 1.91 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 17 | 0.676 | 13 | 1.87 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 18 | 0.674 | 24 | 2.67 | Not shortlisted |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 19 | 0.674 | 19 | 2.11 | Not shortlisted |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 20 | 0.674 | 20 | 2.12 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 21 | 0.671 | 18 | 2.05 | Not shortlisted |
| OpenAI gpt-oss-120b | Off-the-shelf LLM | 22 | 0.669 | 25 | 3.01 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 23 | 0.665 | 22 | 2.44 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 24 | 0.656 | 21 | 2.43 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 25 | 0.655 | 23 | 2.55 | Not shortlisted |
| MiniMax M2.5 | Off-the-shelf LLM | 26 | 0.641 | 28 | 4.39 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 27 | 0.638 | 26 | 3.05 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 28 | 0.633 | 27 | 3.13 | Not shortlisted |
| Cohere Tiny Aya | Adjusted LLM | 29 | 0.437 | 29 | 10.36 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 30 | 0.331 | 30 | 11.10 | Not shortlisted |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.676 | — | 1.83 | Same model family as OpenAI GPT-5.6 Luna |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.690 | — | 1.78 | Same model family as OpenAI GPT-5.6 Luna |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | — | 0.686 | — | 1.99 | Same model family as xAI Grok 4.20 (non-reasoning) |
Every model we got a translation from, as the shortlist compared it: off the shelf, without context in the request, or adjusted, with it. A model that did not make the shortlist is shown adjusted, unless off the shelf it scores higher by 0.03 or more on COMET-Kiwi. Ranks among 30 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 automatic scores rank models within Javanese and are not a requirement result. COMET-Kiwi alone brought in the three solutions that meet tag handling and formatting, on z.ai GLM-5, Anthropic Claude Haiku 4.5 and Cohere Command A Translate. MetricX alone brought in the solution on Anthropic Claude Sonnet 4.6, the only one that meets tone of voice. Nearest the cut are Translated Lara 2, 10th on COMET-Kiwi at 0.705 against 0.706 in 8th place, and OpenAI GPT-5.6 Luna, 9th on MetricX, one place behind Google Gemma 4.
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.14 | Open-weight LLMs |
| OpenAI GPT-5.6 Luna | $0.41 | LLMs |
| MiniMax M2.5 | $0.48 | LLMs |
| Mistral Large 3 | $0.74 | LLMs |
| DeepSeek V3.2 | $0.88 | LLMs |
| Mistral Medium 3.1 | $0.92 | LLMs |
| Alibaba Qwen3-235B-A22B | $1.08 | Open-weight LLMs |
| xAI Grok 4.20 (non-reasoning) | $1.1 | LLMs |
| xAI Grok 4.20 (reasoning) | $1.1 | LLMs |
| MoonshotAI Kimi K2.5 | $1.19 | LLMs |
| z.ai GLM-5 | $1.84 | LLMs |
| Anthropic Claude Haiku 4.5 | $2.14 | LLMs |
| OpenAI GPT-4.1 | $2.7 | LLMs |
| Google Gemini 3.5 Flash | $3.34 | LLMs |
| AI21 Jamba Large 1.7 | $3.6 | LLMs |
| Yandex Cloud Translate v2 | $3.94 | MT engines |
| OpenAI GPT-5.6 Terra | $4.07 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.26 | LLMs |
| Anthropic Claude Sonnet 5 | $5.95 | LLMs |
| Anthropic Claude Sonnet 4.6 | $6.55 | LLMs |
| Alibaba Machine Translation API | $7 | MT engines |
| OpenAI GPT-5.6 Sol | $7.01 | LLMs |
| OpenAI GPT-5.5 | $10.04 | LLMs |
| Anthropic Claude Opus 4.8 | $14.71 | LLMs |
| Google Cloud Advanced Translation API | $20 | MT engines |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $29.57 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Sonnet 5 | 0.496 | LLMs |
| Anthropic Claude Fable 5 | 0.494 | LLMs |
| Anthropic Claude Opus 4.8 | 0.49 | LLMs |
| Mistral Medium 3.1 | 0.403 | LLMs |
| Mistral Large 3 | 0.393 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.37 | LLMs |
| AI21 Jamba Large 1.7 | 0.368 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.36 | LLMs |
| MoonshotAI Kimi K2.5 | 0.336 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.332 | Open-weight LLMs |
| z.ai GLM-5 | 0.329 | LLMs |
| DeepSeek V3.2 | 0.326 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.315 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.314 | LLMs |
| Google Gemini 3.5 Flash | 0.297 | LLMs |
| Google Gemma 4 | 0.297 | Open-weight LLMs |
| Google Gemini 3.1 Pro Preview | 0.295 | LLMs |
| OpenAI GPT-5.6 Terra | 0.282 | LLMs |
| OpenAI GPT-5.6 Sol | 0.281 | LLMs |
| OpenAI GPT-5.6 Luna | 0.28 | LLMs |
| OpenAI GPT-5.5 | 0.274 | LLMs |
| MiniMax M2.5 | 0.273 | LLMs |
| OpenAI GPT-4.1 | 0.272 | LLMs |
| Cohere Command A Translate | 0.264 | LLMs |
| Cohere Tiny Aya | 0.207 | LLMs |
Over the 25 models we count tokens for, a translation into Javanese has a median 0.32 tokens per character, 1.4 times the tokens per character of English text. Tokens alone do not set the cost. Off the shelf, translations from Anthropic Claude Haiku 4.5 and Anthropic Claude Sonnet 4.6 have 0.36 and 0.37 tokens per character, and cost $2.14 and $6.55 per million characters. The shortlisted solution on Anthropic Claude Sonnet 4.6, adjusted for the Standard requirements, costs $7.35 per million characters.
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Google Cloud Advanced Translation API | 0.1 | MT engines |
| ModernMT Enterprise Edition | 0.4 | MT engines |
| DeepL Next Gen | 0.8 | LLM-based MT engines |
| Translated Lara 2 | 1.4 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 2 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 2.4 | MT engines |
| Cohere Tiny Aya | 7.5 | LLMs |
| OpenAI GPT-5.6 Luna | 18 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 20 | LLMs |
| OpenAI GPT-4.1 | 23 | LLMs |
| Alibaba Machine Translation API | 25 | MT engines |
| Anthropic Claude Haiku 4.5 | 33 | LLMs |
| Cohere Command A Translate | 36 | LLMs |
| Swiss AI Apertus-8B-2509 | 39 | Open-weight LLMs |
| OpenAI GPT-5.6 Terra | 40 | LLMs |
| MiniMax M2.5 | 41 | LLMs |
| Mistral Medium 3.1 | 49 | LLMs |
| xAI Grok 4.20 (reasoning) | 50 | LLMs |
| AI21 Jamba Large 1.7 | 51 | LLMs |
| Google Gemini 3.5 Flash | 58 | LLMs |
| Google Gemini 3.1 Pro Preview | 67 | LLMs |
| Anthropic Claude Sonnet 5 | 71 | LLMs |
| Anthropic Claude Fable 5 | 77 | LLMs |
| Anthropic Claude Sonnet 4.6 | 80 | LLMs |
| DeepSeek V3.2 | 83 | LLMs |
| Mistral Large 3 | 84 | LLMs |
| Anthropic Claude Opus 4.8 | 93 | LLMs |
| OpenAI GPT-5.5 | 98 | LLMs |
| OpenAI GPT-5.6 Sol | 106 | LLMs |
| Alibaba Qwen3-235B-A22B | 136 | Open-weight LLMs |
| Google Gemma 4 | 175 | Open-weight LLMs |
| z.ai GLM-5 | 183 | LLMs |
| MoonshotAI Kimi K2.5 | 283 | 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 Javanese.
Of the 5 real-time models in Javanese, 2 are MT engines, 2 are LLM-based MT engines and 1, OpenAI gpt-oss-120b, is an open-weight LLM. Only DeepL Next Gen among them is on the shortlist, and the solution on it misses formatting. Translations from DeepL Next Gen have 4.26% of the sentences we check in the formal register and keep 95.7% of the tags, against a bar of 100% for both. After the six fastest models, five of them real-time, the next fastest translations are from Cohere Tiny Aya, at 7.5 seconds per 1,000 words.
Only the solution on Anthropic Claude Sonnet 4.6 has the formal register in every sentence we check. In translations from the other 12 shortlisted solutions, 2.1% to 97.9% of the sentences we check are in the formal register. In the gender neutrality results shown for information, the solution on Anthropic Claude Sonnet 4.6 has 98.1% of the sentences about people gender-neutral, and the other 12 shortlisted solutions 100%.
Leave feedback