The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Welsh 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 Welsh. |
| Tone of voice | The formal register (chi) 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 Welsh, 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.
Four solutions meet every Standard requirement in Welsh. The two-step solution (translation + post-edit) on Google Gemini 3.1 Pro Preview has the fewest total errors, 15.3 per 1,000 words, at the lowest cost, $5.23 per million characters, so Optimize for selects it for quality and cost. For latency it selects the one-step solution on Anthropic Claude Fable 5, at 91 seconds per 1,000 words against 97, which costs $32.76 per million characters and has 20.2 total errors per 1,000 words.
A cost cap below $5.23 per million characters or a time limit below 91 seconds per 1,000 words removes every solution that meets the Standard requirements in Welsh. A cap below $15.37 per million characters leaves only the two-step solution on Google Gemini 3.1 Pro Preview. A limit of 96 seconds per 1,000 words or less leaves only the one-step solution on Anthropic Claude Fable 5, and so does a limit of one step, at $32.76 per million characters.
For the Advanced requirements, the gender-neutrality agent works with the two-step solutions on OpenAI GPT-5.5 and Google Gemini 3.1 Pro Preview. On OpenAI GPT-5.5 it raises the cost from $43.16 to $48.52 per million characters and the time from 222 to 239 seconds per 1,000 words, and total errors fall from 15.9 to 8.9 per 1,000 words. On Google Gemini 3.1 Pro Preview the cost goes from $5.23 to $4.87 per million characters, the time from 97 to 108 seconds and total errors from 15.3 to 15.8, both per 1,000 words.
OpenAI GPT-5.5
two-step
Google Gemini 3.1 Pro Preview
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Google Gemini 3.1 Pro Preview
two-step
Google Gemini 3.1 Pro Preview
two-step
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
two-step with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
two-step with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
two-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 | ★★ ☆☆ | one-step | Google Gemini 3.1 Pro Preview | 16.0 | $5.23 | 97 |
| Standard | Tier 1: simple solution + backup | ★ ☆☆☆ | one-step | Anthropic Claude Fable 5 | 20.2 | $32.76 | 91 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | with the gender-neutrality agent | Google Gemini 3.1 Pro Preview | 15.8 | $4.87 | 108 |
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.
Under the Basic requirements, Optimize for selects the one-step solution on Google Gemini 3.1 Pro Preview for cost, Anthropic Claude Fable 5 for latency and the two-step solution (translation + post-edit) on OpenAI GPT-5.5 for quality. Tag handling, a Standard requirement, removes the first, and its two-step version takes its place. The Advanced requirements leave the two best solutions, with the gender-neutrality agent, and Optimize for selects OpenAI GPT-5.5 for quality and Google Gemini 3.1 Pro Preview for cost and latency.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| OpenAI GPT-5.6 Luna | 0.46 | 27 | Misses a Standard requirement |
| xAI Grok 4.20 (reasoning) | 1.3 | 45.8 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 3.24 | 47.1 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.92 | 20.4 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 5.23 | 12.9 | Misses a Standard requirement |
| Anthropic Claude Sonnet 5 | 6.54 | 50.1 | Misses a Standard requirement |
| Microsoft Translator | 10 | 60 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 11.64 | 19.1 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 44.2 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 32.76 | 17.8 | Meets every Standard requirement |
Among the one-step solutions in Welsh, weighted errors do not fall steadily as the price rises. The five that cost more than the one on Google Gemini 3.1 Pro Preview all have more weighted errors per 1,000 words than its 12.9, from 17.8 for Anthropic Claude Fable 5 at $32.76 per million characters to 60.0 for Microsoft Translator at $10. The four cheaper ones, all LLMs from $0.46 to $3.92 per million characters, have 20.4 to 47.1 weighted errors per 1,000 words.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 44.2 |
| Gender neutrality | 2.5 |
| Tone of voice | 0 |
| Tag handling | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
Two one-step solutions come close on tag handling in Welsh: the ones on Google Gemini 3.1 Pro Preview and xAI Grok 4.20 (reasoning) have 99.6% of the tags kept, against 96.5% and 94.5% for the ones on Microsoft Translator and DeepL Next Gen. As two-step solutions (translation + post-edit), the one on Google Gemini 3.1 Pro Preview has 100% of the tags kept. The one on OpenAI GPT-5.5 meets general translation quality with one major error against four, and the one on Google Gemini 3.5 Flash with 19.4 weighted errors per 1,000 words.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Welsh | Class |
|---|---|---|
| OpenAI gpt-oss-120b | 189.5 | Open-weight LLMs |
| Microsoft Translator | 60 | MT engines |
| Anthropic Claude Sonnet 5 | 50.1 | LLMs |
| OpenAI GPT-4.1 | 47.1 | LLMs |
| xAI Grok 4.20 (reasoning) | 45.8 | LLMs |
| DeepL Next Gen | 44.2 | LLM-based MT engines |
| OpenAI GPT-5.6 Luna | 27 | LLMs |
| Google Gemini 3.5 Flash | 20.4 | LLMs |
| OpenAI GPT-5.5 | 19.1 | LLMs |
| Anthropic Claude Fable 5 | 17.8 | LLMs |
| Google Gemini 3.1 Pro Preview | 12.9 | LLMs |
Apart from the open-weight OpenAI gpt-oss-120b, the one-step solutions on the eight shortlisted LLMs have 12.9 to 50.1 weighted errors per 1,000 words in Welsh, and they include the two that meet general translation quality. Translations from the LLM-based MT engine DeepL Next Gen have 44.2 weighted errors per 1,000 words, inside that range, and from the MT engine Microsoft Translator 60.0. The two shortlisted LLMs that do not meet our language-support criteria are Anthropic Claude Sonnet 5 and OpenAI gpt-oss-120b.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| z.ai GLM-5 | 0.747 | 9 | 2.80 | 19 | No | Not shortlisted |
| DeepSeek V3.2 | 0.747 | 10 | 2.63 | 9 | No | Not shortlisted |
| Anthropic Claude Opus 4.8 | 0.745 | 12 | 2.87 | 21 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.745 | 13 | 2.64 | 11 | No | Not shortlisted |
| Mistral Large 3 | 0.743 | 14 | 2.67 | 13 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.742 | 15 | 2.78 | 17 | No | Not shortlisted |
| Cohere Command A Translate | 0.740 | 17 | 2.64 | 12 | No | Not shortlisted |
| Translated Lara 2 | 0.738 | 18 | 2.84 | 20 | No | Not shortlisted |
| Google Translation LLM | 0.738 | 19 | 3.07 | 23 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.738 | 20 | 2.69 | 14 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.735 | 22 | 2.72 | 15 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.734 | 23 | 3.00 | 22 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.720 | 24 | 3.08 | 24 | No | Not shortlisted |
| MiniMax M2.5 | 0.718 | 25 | 3.09 | 25 | No | Not shortlisted |
| Amazon Translate | 0.716 | 26 | 3.30 | 27 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.715 | 27 | 3.68 | 29 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.703 | 28 | 3.20 | 26 | No | Not shortlisted |
| Google Gemma 4 | 0.696 | 29 | 3.63 | 28 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.688 | 30 | 3.82 | 30 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.668 | 31 | 4.20 | 32 | No | Not shortlisted |
| Cohere Tiny Aya | 0.667 | 32 | 4.21 | 33 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.648 | 33 | 3.83 | 31 | No | Not shortlisted |
| OpenAI GPT-5.6 Luna | 0.757 | 1 | 2.50 | 2 | Yes | Top 8 by both metrics |
| OpenAI GPT-4.1 | 0.756 | 2 | 2.53 | 4 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.5 | 0.755 | 3 | 2.46 | 1 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.751 | 4 | 2.80 | 18 | Yes | Top 8 by COMET-Kiwi |
| xAI Grok 4.20 (reasoning) | 0.751 | 5 | 2.51 | 3 | Yes | Top 8 by both metrics |
| Google Gemini 3.5 Flash | 0.750 | 6 | 2.63 | 8 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.750 | 7 | 2.57 | 5 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | 0.749 | 8 | 2.63 | 10 | Yes | Top 8 by COMET-Kiwi |
| DeepL Next Gen | 0.747 | 11 | 2.59 | 6 | Yes | Top 8 by MetricX |
| Microsoft Translator | 0.741 | 16 | 2.73 | 16 | Yes | Real-time place |
| OpenAI gpt-oss-120b | 0.735 | 21 | 2.61 | 7 | Yes | Top 8 by MetricX |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| OpenAI GPT-5.6 Luna | Adjusted LLM | 1 | 0.757 | 2 | 2.50 | Top 8 by both metrics |
| OpenAI GPT-4.1 | Adjusted LLM | 2 | 0.756 | 4 | 2.53 | Top 8 by both metrics |
| OpenAI GPT-5.5 | Adjusted LLM | 3 | 0.755 | 1 | 2.46 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 4 | 0.751 | 18 | 2.80 | Top 8 by COMET-Kiwi |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 5 | 0.751 | 3 | 2.51 | Top 8 by both metrics |
| Google Gemini 3.5 Flash | Adjusted LLM | 6 | 0.750 | 8 | 2.63 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 7 | 0.750 | 5 | 2.57 | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 8 | 0.749 | 10 | 2.63 | Top 8 by COMET-Kiwi |
| z.ai GLM-5 | Adjusted LLM | 9 | 0.747 | 19 | 2.80 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 10 | 0.747 | 9 | 2.63 | Not shortlisted |
| DeepL Next Gen | Off-the-shelf NMT | 11 | 0.747 | 6 | 2.59 | Top 8 by MetricX |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 12 | 0.745 | 21 | 2.87 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 13 | 0.745 | 11 | 2.64 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 14 | 0.743 | 13 | 2.67 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 15 | 0.742 | 17 | 2.78 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 16 | 0.741 | 16 | 2.73 | Real-time place |
| Cohere Command A Translate | Adjusted LLM | 17 | 0.740 | 12 | 2.64 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 18 | 0.738 | 20 | 2.84 | Not shortlisted |
| Google Translation LLM | Adjusted NMT | 19 | 0.738 | 23 | 3.07 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 20 | 0.738 | 14 | 2.69 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 21 | 0.735 | 7 | 2.61 | Top 8 by MetricX |
| Mistral Medium 3.1 | Adjusted LLM | 22 | 0.735 | 15 | 2.72 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 23 | 0.734 | 22 | 3.00 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 24 | 0.720 | 24 | 3.08 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 25 | 0.718 | 25 | 3.09 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 26 | 0.716 | 27 | 3.30 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 27 | 0.715 | 29 | 3.68 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 28 | 0.703 | 26 | 3.20 | Not shortlisted |
| Google Gemma 4 | Adjusted LLM | 29 | 0.696 | 28 | 3.63 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 30 | 0.688 | 30 | 3.82 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 31 | 0.668 | 32 | 4.20 | Not shortlisted |
| Cohere Tiny Aya | Adjusted LLM | 32 | 0.667 | 33 | 4.21 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 33 | 0.648 | 31 | 3.83 | Not shortlisted |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.757 | — | 2.42 | Same model family as OpenAI GPT-5.6 Luna |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.749 | — | 2.46 | Same model family as OpenAI GPT-5.6 Luna |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.746 | — | 2.48 | 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 33 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 solutions differently in Welsh. The one-step solution on OpenAI GPT-5.6 Luna, first on COMET-Kiwi, has 27.0 weighted errors per 1,000 words, and the one on Google Gemini 3.1 Pro Preview, 7th on COMET-Kiwi and 5th on MetricX, has the fewest, 12.9. Anthropic Claude Fable 5 is on the list through COMET-Kiwi alone, 18th on MetricX, and its solution is the only one-step solution that meets every Standard requirement. Nearest the cut are z.ai GLM-5, 9th on COMET-Kiwi, and DeepSeek V3.2, 9th on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | Google Gemini 3.1 Pro Preview, one-step LLM adjusted for Standard requirements | $5.23 |
| Standard | Google Gemini 3.1 Pro Preview, two-step | $5.23 |
| Advanced | Google Gemini 3.1 Pro Preview, two-step with the gender-neutrality agent | $4.87 |
| Model | US dollars per million characters | Class |
|---|---|---|
| OpenAI GPT-5.6 Luna | $0.46 | LLMs |
| xAI Grok 4.20 (reasoning) | $1.3 | LLMs |
| OpenAI GPT-4.1 | $3.24 | LLMs |
| Google Gemini 3.5 Flash | $3.92 | LLMs |
| Google Gemini 3.1 Pro Preview | $5.23 | LLMs |
| Anthropic Claude Sonnet 5 | $6.54 | LLMs |
| Microsoft Translator | $10 | MT engines |
| OpenAI GPT-5.5 | $11.64 | LLMs |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $32.76 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Sonnet 5 | 0.555 | LLMs |
| Anthropic Claude Opus 4.8 | 0.554 | LLMs |
| Anthropic Claude Fable 5 | 0.552 | LLMs |
| Mistral Medium 3.1 | 0.454 | LLMs |
| Mistral Large 3 | 0.45 | LLMs |
| AI21 Jamba Large 1.7 | 0.428 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.424 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.423 | LLMs |
| Cohere Command A Translate | 0.408 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.395 | Open-weight LLMs |
| DeepSeek V3.2 | 0.393 | LLMs |
| z.ai GLM-5 | 0.39 | LLMs |
| MoonshotAI Kimi K2.5 | 0.384 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.38 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.379 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.376 | LLMs |
| Google Gemini 3.5 Flash | 0.376 | LLMs |
| Google Gemma 4 | 0.373 | Open-weight LLMs |
| Cohere Tiny Aya | 0.338 | LLMs |
| OpenAI GPT-5.6 Terra | 0.329 | LLMs |
| OpenAI GPT-4.1 | 0.328 | LLMs |
| OpenAI GPT-5.5 | 0.328 | LLMs |
| OpenAI GPT-5.6 Sol | 0.328 | LLMs |
| MiniMax M2.5 | 0.327 | LLMs |
| OpenAI GPT-5.6 Luna | 0.327 | LLMs |
With billing on tokens, the tokenizers produce a median 0.38 tokens per character of Welsh, 1.6 times their count for English. The steps change the cost too. For OpenAI GPT-5.5, the two-step solution (translation + post-edit) costs $43.16 per million characters against $11.64 for the one-step solution, and for Google Gemini 3.5 Flash $15.37 against $3.92. For Google Gemini 3.1 Pro Preview both cost $5.23 per million characters, so the cheapest solution costs the same under the Basic and the Standard requirements.
| 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 | 91 |
| Standard | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | 91 |
| Advanced | Google Gemini 3.1 Pro Preview, two-step with the gender-neutrality agent | 108 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Google Cloud Advanced Translation API | 0.1 | MT engines |
| Microsoft Translator | 0.2 | MT engines |
| ModernMT Enterprise Edition | 0.3 | MT engines |
| DeepL Next Gen | 0.7 | LLM-based MT engines |
| Google Translation LLM | 0.7 | LLM-based MT engines |
| Amazon Translate | 1.4 | MT engines |
| Translated Lara 2 | 1.7 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 2.6 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 4 | MT engines |
| OpenAI GPT-5.6 Luna | 19 | LLMs |
| Cohere Tiny Aya | 21 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 29 | LLMs |
| Alibaba Machine Translation API | 40 | MT engines |
| Anthropic Claude Haiku 4.5 | 41 | LLMs |
| Mistral Medium 3.1 | 48 | LLMs |
| Swiss AI Apertus-8B-2509 | 51 | Open-weight LLMs |
| OpenAI GPT-5.6 Terra | 55 | LLMs |
| MiniMax M2.5 | 56 | LLMs |
| OpenAI GPT-4.1 | 59 | LLMs |
| AI21 Jamba Large 1.7 | 60 | LLMs |
| xAI Grok 4.20 (reasoning) | 64 | LLMs |
| Cohere Command A Translate | 70 | LLMs |
| Mistral Large 3 | 77 | LLMs |
| Google Gemini 3.5 Flash | 80 | LLMs |
| DeepSeek V3.2 | 84 | LLMs |
| Anthropic Claude Sonnet 5 | 85 | LLMs |
| Anthropic Claude Sonnet 4.6 | 85 | LLMs |
| OpenAI GPT-5.6 Sol | 86 | LLMs |
| Anthropic Claude Fable 5 | 91 | LLMs |
| Google Gemini 3.1 Pro Preview | 97 | LLMs |
| Anthropic Claude Opus 4.8 | 110 | LLMs |
| OpenAI GPT-5.5 | 123 | LLMs |
| Alibaba Qwen3-235B-A22B | 158 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 205 | LLMs |
| Google Gemma 4 | 235 | Open-weight LLMs |
| z.ai GLM-5 | 289 | 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 Welsh.
Seven of the eight real-time models in Welsh are MT engines or LLM-based MT engines, and the other is the open-weight OpenAI gpt-oss-120b, whose one-step solution has 189.5 weighted errors per 1,000 words. Among the other shortlisted LLMs, the one-step solutions range from 19 seconds per 1,000 words on OpenAI GPT-5.6 Luna to 123 on OpenAI GPT-5.5. On OpenAI GPT-5.5 the two-step solution (translation + post-edit) takes 222 seconds per 1,000 words, and 239 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 | two-step, with the gender-neutrality agent | 100% | ✓ | $4.87 | 108 |
| OpenAI GPT-5.5 | two-step, with the gender-neutrality agent | 100% | ✓ | $48.52 | 239 |
Tone of voice in Welsh asks for the formal register, with chi. For gender neutrality we check neutral nouns and rephrasing, with no default gender. The Advanced requirements count the two best solutions, the two-step solutions (translation + post-edit) on Google Gemini 3.1 Pro Preview and OpenAI GPT-5.5 with the gender-neutrality agent. In the translations from both, 100% of the sentences about people are gender-neutral. Without the agent, each solution has one major error in general translation quality, and with it none.
Both pairs of automatic metrics are above the threshold in Welsh, and neither is borderline, so the shortlist rests on measurement we can rely on. The three LQA judges agree on a median of 94.4% of the sentences of the 11 shortlisted solutions, so we can rely on their Intento LQA results too.
Leave feedback