The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Thai 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 Thai. |
| Tone of voice | The formal register (ท่าน or neutral forms) 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 Thai, with a column for each requirement you select. Under Optimize for, choose which of them comes first: the one with the fewest total errors, the lowest cost or the shortest time. Click a column heading to sort by it.
The shortlisted solutions on 13 models, with their Intento LQA results. Total errors per 1,000 words add the weighted general translation quality errors (minor 1, major 5, critical 10) to the missed glossary terms, sentences in the wrong register, tag errors and sentences that fail gender neutrality, each counted once. Formatting counts as met or missed and does not add to the total errors. The models that did not make the shortlist have automatic scores only: the shortlist. 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.
Four three-step solutions (translation + LQA + repair) meet every Standard requirement in Thai. The one on OpenAI GPT-5.6 Sol has the fewest total errors, 9.2 per 1,000 words, and the lowest cost, $554 per million characters, so optimizing for quality or cost selects it. Optimizing for latency selects the one on Google Gemini 3.5 Flash, at 165 seconds per 1,000 words against 180 for OpenAI GPT-5.6 Sol. It costs $562 per million characters and has 18.8 total errors per 1,000 words.
The Advanced requirements add gender neutrality, which they count on the two best solutions in Thai, the three-step ones on OpenAI GPT-5.6 Sol and Anthropic Claude Fable 5. In translations from both, 100% of the sentences about people are gender-neutral without the gender-neutrality agent, so their cost, time and errors are the same as for the Standard requirements. Without the solution on Google Gemini 3.5 Flash, the shortest time goes from 165 to 180 seconds per 1,000 words.
OpenAI GPT-5.6 Sol
three-step
Anthropic Claude Opus 4.8
one-step LLM adjusted for Standard requirements
Anthropic Claude Opus 4.8
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
three-step
OpenAI GPT-5.6 Sol
three-step
Google Gemini 3.5 Flash
three-step
OpenAI GPT-5.6 Sol
three-step
OpenAI GPT-5.6 Sol
three-step
OpenAI GPT-5.6 Sol
three-step
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 | ★ ☆×5 | one-step | Anthropic Claude Opus 4.8 | 20.5 | $21.28 | 119 |
| Standard | Tier 2: two or more solutions, none simple | ☆×4 | three-step | OpenAI GPT-5.6 Sol | 9.2 | $554 | 180 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | three-step | OpenAI GPT-5.6 Sol | 9.2 | $554 | 180 |
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 six models with a solution that meets the Basic ones in Thai, Anthropic Claude Opus 4.8 and Anthropic Claude Sonnet 5, both on tag handling. Translations from those solutions have 99.6% to 99.8% of the tags right. For the Basic requirements, optimizing for cost or latency selects the one-step solution on Anthropic Claude Opus 4.8, and for quality the three-step one (translation + LQA + repair) on OpenAI GPT-5.6 Sol. For the Advanced requirements, all three options select OpenAI GPT-5.6 Sol.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.13 | 49.8 | Misses a Standard requirement |
| Mistral Large 3 | 1.35 | 63 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.42 | 28.7 | Misses a Standard requirement |
| z.ai GLM-5 | 3.53 | 43.2 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.48 | 35.9 | Misses a Standard requirement |
| Anthropic Claude Sonnet 5 | 8.28 | 33.9 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 8.46 | 22.4 | Misses a Standard requirement |
| Google Translation LLM | 20.63 | 84.4 | Misses a Standard requirement |
| Anthropic Claude Opus 4.8 | 21.28 | 19.8 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 43.8 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 66.2 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 41.54 | 24.4 | Misses a Standard requirement |
Among the one-step solutions in Thai, the dearest is not the one with the fewest errors. The one on Anthropic Claude Fable 5 costs $41.54 per million characters and has 24.4 weighted errors per 1,000 words, against 19.8 for Anthropic Claude Opus 4.8 at $21.28 and 22.4 for OpenAI GPT-5.6 Sol at $8.46. The ones on the three LLM-based MT engines cost $20.63 to $27.50 per million characters and have 43.8 to 84.4 weighted errors per 1,000 words, against $3.42 and 28.7 for Google Gemini 3.5 Flash.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 39.5 |
| Tone of voice | 0.6 |
| Tag handling | 0 |
| Gender neutrality | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Thai | Class |
|---|---|---|
| Google Translation LLM | 84.4 | LLM-based MT engines |
| DeepL Next Gen | 66.2 | LLM-based MT engines |
| Mistral Large 3 | 63 | LLMs |
| Google Gemma 4 | 49.8 | Open-weight LLMs |
| Translated Lara 2 | 43.8 | LLM-based MT engines |
| z.ai GLM-5 | 43.2 | LLMs |
| Google Gemini 3.1 Pro Preview | 35.9 | LLMs |
| Anthropic Claude Sonnet 5 | 33.9 | LLMs |
| Google Gemini 3.5 Flash | 28.7 | LLMs |
| Anthropic Claude Fable 5 | 24.4 | LLMs |
| OpenAI GPT-5.6 Sol | 22.4 | LLMs |
| Anthropic Claude Opus 4.8 | 19.8 | LLMs |
No translation into Thai: Widn AI Tower Anthill (far too short, wrong language, wrong script), Widn AI Tower Vesuvius (far too short, wrong language, wrong script).
Translations from the eight shortlisted LLMs that meet our language-support criteria in Thai have 19.8 to 49.8 weighted errors per 1,000 words in one step, the most for the open-weight Google Gemma 4. All three shortlisted MT engines are LLM-based: DeepL Next Gen, Translated Lara 2 and Google Translation LLM. Translations from them have 1 to 6 critical errors, against at most one for each of those LLMs. Translations from Mistral Large 3, the one shortlisted LLM outside the criteria, have 63.0 weighted errors per 1,000 words.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Swiss AI Apertus-8B-2509 | 0.784 | 9 | 1.30 | 16 | No | Not shortlisted |
| OpenAI GPT-5.5 | 0.783 | 12 | 1.35 | 17 | No | Not shortlisted |
| Meta Llama 4 Maverick | 0.783 | 13 | 1.27 | 15 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.780 | 15 | 1.25 | 14 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.779 | 16 | 1.24 | 13 | No | Not shortlisted |
| xAI Grok 4.20 (non-reasoning) | 0.778 | 17 | 1.40 | 20 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.778 | 18 | 1.23 | 12 | No | Not shortlisted |
| DeepSeek V3.2 | 0.773 | 19 | 1.37 | 18 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.770 | 22 | 1.47 | 22 | No | Not shortlisted |
| MiniMax M2.5 | 0.768 | 23 | 1.39 | 19 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.764 | 24 | 1.49 | 23 | No | Not shortlisted |
| Microsoft Translator | 0.763 | 25 | 1.47 | 21 | No | Not shortlisted |
| Cohere Command A Translate | 0.762 | 26 | 1.58 | 26 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.759 | 27 | 2.08 | 32 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.755 | 28 | 1.54 | 24 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.749 | 29 | 1.58 | 25 | No | Not shortlisted |
| Amazon Translate | 0.742 | 30 | 1.84 | 29 | No | Not shortlisted |
| Naver Papago NMT Commercial | 0.739 | 31 | 1.89 | 30 | No | Not shortlisted |
| Tencent Cloud TMT | 0.735 | 32 | 1.92 | 31 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.709 | 33 | 2.30 | 33 | No | Not shortlisted |
| Baidu Translate API | 0.691 | 34 | 2.84 | 34 | No | Not shortlisted |
| Cohere Tiny Aya | 0.689 | 35 | 3.00 | 35 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.656 | 36 | 3.20 | 36 | No | Not shortlisted |
| AppTek NMT | 0.605 | 37 | 4.51 | 38 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.581 | 38 | 4.45 | 37 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.563 | 39 | 7.29 | 39 | No | Not shortlisted |
| Google Gemini 3.5 Flash | 0.792 | 1 | 1.06 | 1 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | 0.791 | 2 | 1.15 | 3 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | 0.791 | 3 | 1.16 | 5 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.790 | 4 | 1.15 | 4 | Yes | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | 0.789 | 5 | 1.19 | 8 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.787 | 6 | 1.19 | 9 | Yes | Top 8 by COMET-Kiwi |
| OpenAI GPT-4.1 | 0.786 | 7 | 1.20 | 10 | Yes | Top 8 by COMET-Kiwi |
| z.ai GLM-5 | 0.786 | 8 | 1.18 | 6 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.784 | 10 | 1.15 | 2 | Yes | Top 8 by MetricX |
| DeepL Next Gen | 0.784 | 11 | 1.21 | 11 | Yes | Real-time place |
| Mistral Large 3 | 0.781 | 14 | 1.19 | 7 | Yes | Top 8 by MetricX |
| Translated Lara 2 | 0.773 | 20 | 1.67 | 28 | Yes | Real-time place |
| Google Translation LLM | 0.771 | 21 | 1.59 | 27 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| Google Gemini 3.5 Flash | Adjusted LLM | 1 | 0.792 | 1 | 1.06 | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 2 | 0.791 | 3 | 1.15 | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | Adjusted LLM | 3 | 0.791 | 5 | 1.16 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 4 | 0.790 | 4 | 1.15 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 5 | 0.789 | 8 | 1.19 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 6 | 0.787 | 9 | 1.19 | Top 8 by COMET-Kiwi |
| OpenAI GPT-4.1 | Adjusted LLM | 7 | 0.786 | 10 | 1.20 | Top 8 by COMET-Kiwi |
| z.ai GLM-5 | Adjusted LLM | 8 | 0.786 | 6 | 1.18 | Top 8 by both metrics |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 9 | 0.784 | 16 | 1.30 | Not shortlisted |
| Google Gemma 4 | Adjusted LLM | 10 | 0.784 | 2 | 1.15 | Top 8 by MetricX |
| DeepL Next Gen | Off-the-shelf NMT | 11 | 0.784 | 11 | 1.21 | Real-time place |
| OpenAI GPT-5.5 | Adjusted LLM | 12 | 0.783 | 17 | 1.35 | Not shortlisted |
| Meta Llama 4 Maverick | Adjusted LLM | 13 | 0.783 | 15 | 1.27 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 14 | 0.781 | 7 | 1.19 | Top 8 by MetricX |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 15 | 0.780 | 14 | 1.25 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 16 | 0.779 | 13 | 1.24 | Not shortlisted |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | 17 | 0.778 | 20 | 1.40 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 18 | 0.778 | 12 | 1.23 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 19 | 0.773 | 18 | 1.37 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 20 | 0.773 | 28 | 1.67 | Real-time place |
| Google Translation LLM | Adjusted NMT | 21 | 0.771 | 27 | 1.59 | Real-time place |
| OpenAI gpt-oss-120b | Adjusted LLM | 22 | 0.770 | 22 | 1.47 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 23 | 0.768 | 19 | 1.39 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 24 | 0.764 | 23 | 1.49 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 25 | 0.763 | 21 | 1.47 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 26 | 0.762 | 26 | 1.58 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 27 | 0.759 | 32 | 2.08 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 28 | 0.755 | 24 | 1.54 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 29 | 0.749 | 25 | 1.58 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 30 | 0.742 | 29 | 1.84 | Not shortlisted |
| Naver Papago NMT Commercial | Off-the-shelf NMT | 31 | 0.739 | 30 | 1.89 | Not shortlisted |
| Tencent Cloud TMT | Off-the-shelf NMT | 32 | 0.735 | 31 | 1.92 | Not shortlisted |
| SYSTRAN PNMT API | Off-the-shelf NMT | 33 | 0.709 | 33 | 2.30 | Not shortlisted |
| Baidu Translate API | Off-the-shelf NMT | 34 | 0.691 | 34 | 2.84 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 35 | 0.689 | 35 | 3.00 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 36 | 0.656 | 36 | 3.20 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 37 | 0.605 | 38 | 4.51 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 38 | 0.581 | 37 | 4.45 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 39 | 0.563 | 39 | 7.29 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.787 | — | 1.17 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.780 | — | 1.25 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | — | 0.771 | — | 1.52 | 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 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 one-step solution on Google Gemini 3.5 Flash, first on both metrics, has 28.7 weighted errors per 1,000 words by Intento LQA, against 19.8 for Anthropic Claude Opus 4.8, fifth on COMET-Kiwi and eighth on MetricX. Anthropic Claude Fable 5 is on the shortlist through COMET-Kiwi alone, sixth against ninth on MetricX, and the three-step solution (translation + LQA + repair) on it meets every Advanced requirement. Nearest the cut in the tail are Swiss AI Apertus-8B-2509, ninth on COMET-Kiwi, and Mistral Medium 3.1, 12th on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | Anthropic Claude Opus 4.8, one-step LLM adjusted for Standard requirements | $21.28 |
| Standard | OpenAI GPT-5.6 Sol, three-step | $554 |
| Advanced | OpenAI GPT-5.6 Sol, three-step | $554 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.13 | Open-weight LLMs |
| Mistral Large 3 | $1.35 | LLMs |
| OpenAI GPT-4.1 | $3.37 | LLMs |
| Google Gemini 3.5 Flash | $3.42 | LLMs |
| z.ai GLM-5 | $3.53 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.48 | LLMs |
| Anthropic Claude Sonnet 5 | $8.28 | LLMs |
| OpenAI GPT-5.6 Sol | $8.46 | LLMs |
| Google Translation LLM | $20.63 | LLM-based MT engines |
| Anthropic Claude Opus 4.8 | $21.28 | LLMs |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $41.54 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Mistral Medium 3.1 | 0.839 | LLMs |
| Mistral Large 3 | 0.834 | LLMs |
| AI21 Jamba Large 1.7 | 0.82 | LLMs |
| Cohere Command A Translate | 0.819 | LLMs |
| Anthropic Claude Fable 5 | 0.764 | LLMs |
| Anthropic Claude Sonnet 5 | 0.76 | LLMs |
| Anthropic Claude Opus 4.8 | 0.758 | LLMs |
| z.ai GLM-5 | 0.734 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.724 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.714 | LLMs |
| MiniMax M2.5 | 0.639 | LLMs |
| MoonshotAI Kimi K2.5 | 0.63 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.477 | Open-weight LLMs |
| OpenAI GPT-5.6 Luna | 0.377 | LLMs |
| OpenAI GPT-5.6 Sol | 0.376 | LLMs |
| OpenAI GPT-5.6 Terra | 0.376 | LLMs |
| OpenAI GPT-5.5 | 0.374 | LLMs |
| OpenAI GPT-4.1 | 0.371 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.37 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.368 | LLMs |
| DeepSeek V3.2 | 0.345 | LLMs |
| Google Gemini 3.5 Flash | 0.322 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.321 | LLMs |
| Google Gemma 4 | 0.318 | Open-weight LLMs |
| Cohere Tiny Aya | 0.296 | LLMs |
The tokenizers of the three Anthropic Claude models on the shortlist produce 0.758 to 0.764 tokens per character in Thai, against 0.318 to 0.322 for the three Google LLMs, so the same translation takes more tokens on the Anthropic models. On OpenAI GPT-5.6 Sol the cost follows the steps. The one-step solution costs $8.46 per million characters, and the three-step one (translation + LQA + repair), the cheapest solution that meets every Standard requirement, costs $554.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | Anthropic Claude Opus 4.8, one-step LLM adjusted for Standard requirements | 119 |
| Standard | Google Gemini 3.5 Flash, three-step | 165 |
| Advanced | OpenAI GPT-5.6 Sol, three-step | 180 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.2 | MT engines |
| Google Cloud Advanced Translation API | 0.2 | MT engines |
| ModernMT Enterprise Edition | 0.4 | MT engines |
| Google Translation LLM | 0.5 | LLM-based MT engines |
| DeepL Next Gen | 1.3 | LLM-based MT engines |
| Amazon Translate | 1.3 | MT engines |
| SYSTRAN PNMT API | 1.5 | MT engines |
| Translated Lara 2 | 1.8 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 2.1 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 2.5 | MT engines |
| Baidu Translate API | 11 | MT engines |
| Cohere Tiny Aya | 16 | LLMs |
| OpenAI GPT-5.6 Luna | 16 | LLMs |
| OpenAI GPT-4.1 | 18 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 18 | LLMs |
| Alibaba Machine Translation API | 20 | MT engines |
| OpenAI GPT-5.6 Terra | 26 | LLMs |
| DeepSeek V3.2 | 34 | LLMs |
| Meta Llama 4 Maverick | 40 | Open-weight LLMs |
| Naver Papago NMT Commercial | 48 | MT engines |
| xAI Grok 4.20 (reasoning) | 51 | LLMs |
| OpenAI GPT-5.5 | 53 | LLMs |
| Anthropic Claude Haiku 4.5 | 53 | LLMs |
| Mistral Medium 3.1 | 54 | LLMs |
| Tencent Cloud TMT | 56 | MT engines |
| Google Gemini 3.5 Flash | 58 | LLMs |
| OpenAI GPT-5.6 Sol | 58 | LLMs |
| MiniMax M2.5 | 66 | LLMs |
| Cohere Command A Translate | 70 | LLMs |
| Google Gemini 3.1 Pro Preview | 77 | LLMs |
| Mistral Large 3 | 78 | LLMs |
| Anthropic Claude Sonnet 5 | 86 | LLMs |
| Swiss AI Apertus-8B-2509 | 91 | Open-weight LLMs |
| AppTek NMT | 95 | MT engines |
| Anthropic Claude Fable 5 | 103 | LLMs |
| AI21 Jamba Large 1.7 | 109 | LLMs |
| Alibaba Qwen3-235B-A22B | 111 | Open-weight LLMs |
| Anthropic Claude Sonnet 4.6 | 112 | LLMs |
| Anthropic Claude Opus 4.8 | 119 | LLMs |
| Google Gemma 4 | 133 | Open-weight LLMs |
| z.ai GLM-5 | 266 | LLMs |
| MoonshotAI Kimi K2.5 | 278 | 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 Thai.
The nine real-time models (<2 sec) in Thai are five MT engines, three LLM-based MT engines and the open-weight OpenAI gpt-oss-120b. Translations from the three on the shortlist have 80.9% to 89.4% of the sentences in the tone of voice. Only those from Translated Lara 2 have every tag right. One-step solutions on LLMs with a conclusive Intento LQA result take 58 to 266 seconds per 1,000 words. The fastest solutions for the Basic, Standard and Advanced requirements take 119, 165 and 180 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 |
|---|---|---|---|---|---|
| OpenAI GPT-5.6 Sol | three-step | 100% | ✓ | $554 | 180 |
| Anthropic Claude Fable 5 | three-step | 100% | ✓ | $631 | 296 |
For Thai we check the formal register, with ท่าน or neutral forms, and gender-neutral wording through เรา (we), neutral role nouns with ผู้- or นัก-, and omission or rephrasing, with no default gender. Other options in Thai include gendered first-person forms such as ผม or ดิฉัน, gendered particles and gendered role nouns. In translations from both of the two best solutions, 100% of the sentences about people are gender-neutral without the gender-neutrality agent.
For the 13 shortlisted solutions in Thai, the median share of sentences on which the three LQA judges agree is 95.5%. The agreement of both pairs of automatic metrics is not borderline: the lower end of each range is above its floor.
Leave feedback