The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Basque 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 Basque. |
| Tone of voice | The formal register (zu (zuka)) 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 Basque, with a column for each requirement you select. Under Optimize for, choose which of them comes first: the one with the fewest total errors, the lowest cost or the shortest time. Click a column heading to sort by it.
The shortlisted solutions on 12 models, with their Intento LQA results. Total errors per 1,000 words add the weighted general translation quality errors (minor 1, major 5, critical 10) to the missed glossary terms, sentences in the wrong register, tag errors and sentences that fail gender neutrality, each counted once. Formatting counts as met or missed and does not add to the total errors. The models that did not make the shortlist have automatic scores only: the shortlist. Inconclusive: the three LQA judges agree on fewer than 90% of the sentences of the solution, so it has no general translation quality result (09-02). Gender neutrality: the Advanced requirements count the two best solutions of the language, with the gender-neutrality agent where they need it. The results of the other solutions are shown for information.
Two three-step solutions (translation + LQA + repair) meet every Standard requirement in Basque, and the one on OpenAI GPT-5.6 Sol has the fewest errors, the lowest cost and the shortest time. It has 11.9 total errors per 1,000 words against 13.5 for the one on Google Gemini 3.1 Pro Preview. It costs $574 against $577 per million characters and takes 209 against 273 seconds per 1,000 words. So optimizing for quality, cost or latency selects the same solution.
Under a cost cap of at most $573 per million characters, a time limit of at most 208 seconds per 1,000 words or a limit of two steps, no solution meets even the Basic requirements in Basque. The three-step solution on Google Gemini 3.5 Flash is faster than the two that meet them, at 172 seconds per 1,000 words and $585 per million characters. It misses general translation quality on major errors alone, with 4 major errors and 18.1 weighted errors per 1,000 words.
The Advanced requirements add gender neutrality. Both three-step solutions (translation + LQA + repair) that meet every Standard requirement in Basque have 100% of the sentences about people gender-neutral without the gender-neutrality agent. So the solution on OpenAI GPT-5.6 Sol meets every Advanced requirement with the same 11.9 total errors per 1,000 words, $574 per million characters and 209 seconds per 1,000 words as under the Standard requirements.
OpenAI GPT-5.6 Sol
three-step
OpenAI GPT-5.6 Sol
three-step
OpenAI GPT-5.6 Sol
three-step
OpenAI GPT-5.6 Sol
three-step
OpenAI GPT-5.6 Sol
three-step
OpenAI GPT-5.6 Sol
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 2: two or more solutions, none simple | ☆☆ | three-step | OpenAI GPT-5.6 Sol | 11.9 | $574 | 209 |
| Standard | Tier 2: two or more solutions, none simple | ☆☆ | three-step | OpenAI GPT-5.6 Sol | 11.9 | $574 | 209 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | three-step | OpenAI GPT-5.6 Sol | 11.9 | $574 | 209 |
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.
In Basque two three-step solutions (translation + LQA + repair), on OpenAI GPT-5.6 Sol and Google Gemini 3.1 Pro Preview, meet general translation quality. Both also meet tone of voice, tag handling and formatting, and gender neutrality without the gender-neutrality agent. So neither the Standard nor the Advanced requirements remove a solution, and every level takes three steps. At each level, optimizing for quality, cost or latency selects the solution on OpenAI GPT-5.6 Sol, with 11.9 total errors per 1,000 words.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| z.ai GLM-5 | 2.01 | 96.9 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 3.11 | 87.3 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.64 | 39.2 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.87 | 26.4 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 7.92 | 21.8 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 11.47 | 34.6 | Misses a Standard requirement |
| Anthropic Claude Opus 4.8 | 15.2 | 51.7 | Misses a Standard requirement |
| Google Translation LLM | 21 | 76.1 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 127.2 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 30.3 | 35.9 | Misses a Standard requirement |
Up to $7.92 per million characters, a dearer one-step solution in Basque has fewer weighted errors. The five there with a conclusive Intento LQA result are on LLMs, from the one on z.ai GLM-5, at $2.01 per million characters and 96.9 weighted errors per 1,000 words, to the one on OpenAI GPT-5.6 Sol, at 21.8 weighted errors per 1,000 words. The five such solutions that cost more, $11.47 to $30.30 per million characters, have 34.6 to 127.2 weighted errors per 1,000 words. Two of them are on LLM-based MT engines.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 45.5 |
| Tone of voice | 0 |
| Tag handling | 0 |
| Gender neutrality | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
The one-step solution closest to the bar of general translation quality is on OpenAI GPT-5.6 Sol, with 21.8 weighted errors per 1,000 words and no major or critical error. Two-step solutions (translation + post-edit) on four models have 25.0 to 83.4 weighted errors per 1,000 words. Three-step solutions (translation + LQA + repair) on four models stay under the bar on weighted errors, but those on OpenAI GPT-5.5 and Google Gemini 3.5 Flash have 5 and 4 major errors. Every three-step solution meets tag handling and formatting.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Basque | Class |
|---|---|---|
| DeepL Next Gen | 127.2 | LLM-based MT engines |
| z.ai GLM-5 | 96.9 | LLMs |
| OpenAI GPT-4.1 | 87.3 | LLMs |
| Google Translation LLM | 76.1 | LLM-based MT engines |
| Anthropic Claude Opus 4.8 | 51.7 | LLMs |
| Google Gemini 3.5 Flash | 39.2 | LLMs |
| Anthropic Claude Fable 5 | 35.9 | LLMs |
| OpenAI GPT-5.5 | 34.6 | LLMs |
| Google Gemini 3.1 Pro Preview | 26.4 | LLMs |
| OpenAI GPT-5.6 Sol | 21.8 | LLMs |
The shortlist for Basque holds nine LLMs and three LLM-based MT engines. The four open-weight LLMs and five other MT engines we got translations from are outside it. For one LLM, Anthropic Claude Sonnet 5, we have no conclusive Intento LQA result. Our language-support criteria cover the other eight. In one step, translations from six of them have 21.8 to 51.7 weighted errors per 1,000 words and meet the criteria. Those from OpenAI GPT-4.1 and z.ai GLM-5 have 87.3 and 96.9 weighted errors per 1,000 words, above the 60 of the criteria.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| DeepSeek V3.2 | 0.779 | 11 | 2.43 | 13 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.779 | 12 | 2.40 | 11 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.770 | 15 | 2.57 | 14 | No | Not shortlisted |
| Microsoft Translator | 0.766 | 16 | 2.93 | 19 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.763 | 17 | 2.60 | 15 | No | Not shortlisted |
| Mistral Large 3 | 0.763 | 18 | 3.09 | 20 | No | Not shortlisted |
| Elia Machine Translation System | 0.762 | 19 | 2.70 | 17 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.758 | 20 | 2.73 | 18 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.753 | 21 | 3.17 | 22 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.740 | 22 | 3.16 | 21 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.739 | 23 | 3.18 | 23 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.723 | 24 | 3.69 | 25 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.722 | 25 | 3.92 | 28 | No | Not shortlisted |
| Google Gemma 4 | 0.721 | 26 | 4.21 | 29 | No | Not shortlisted |
| MiniMax M2.5 | 0.718 | 27 | 3.55 | 24 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.707 | 28 | 3.82 | 27 | No | Not shortlisted |
| Cohere Command A Translate | 0.703 | 29 | 3.72 | 26 | No | Not shortlisted |
| Cohere Tiny Aya | 0.643 | 30 | 5.63 | 30 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.605 | 31 | 6.10 | 31 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.545 | 32 | 8.31 | 32 | No | Not shortlisted |
| Google Gemini 3.1 Pro Preview | 0.799 | 1 | 2.03 | 1 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.5 | 0.795 | 2 | 2.13 | 5 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | 0.795 | 3 | 2.04 | 2 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.790 | 4 | 2.07 | 3 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | 0.789 | 5 | 2.36 | 10 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Opus 4.8 | 0.786 | 6 | 2.28 | 7 | Yes | Top 8 by both metrics |
| z.ai GLM-5 | 0.783 | 7 | 2.34 | 9 | Yes | Top 8 by COMET-Kiwi |
| OpenAI GPT-4.1 | 0.782 | 8 | 2.32 | 8 | Yes | Top 8 by both metrics |
| DeepL Next Gen | 0.782 | 9 | 2.11 | 4 | Yes | Top 8 by MetricX |
| Google Gemini 3.5 Flash | 0.779 | 10 | 2.27 | 6 | Yes | Top 8 by MetricX |
| Google Translation LLM | 0.775 | 13 | 2.41 | 12 | Yes | Real-time place |
| Translated Lara 2 | 0.775 | 14 | 2.69 | 16 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 1 | 0.799 | 1 | 2.03 | Top 8 by both metrics |
| OpenAI GPT-5.5 | Adjusted LLM | 2 | 0.795 | 5 | 2.13 | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | Adjusted LLM | 3 | 0.795 | 2 | 2.04 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 4 | 0.790 | 3 | 2.07 | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 5 | 0.789 | 10 | 2.36 | Top 8 by COMET-Kiwi |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 6 | 0.786 | 7 | 2.28 | Top 8 by both metrics |
| z.ai GLM-5 | Adjusted LLM | 7 | 0.783 | 9 | 2.34 | Top 8 by COMET-Kiwi |
| OpenAI GPT-4.1 | Adjusted LLM | 8 | 0.782 | 8 | 2.32 | Top 8 by both metrics |
| DeepL Next Gen | Off-the-shelf NMT | 9 | 0.782 | 4 | 2.11 | Top 8 by MetricX |
| Google Gemini 3.5 Flash | Adjusted LLM | 10 | 0.779 | 6 | 2.27 | Top 8 by MetricX |
| DeepSeek V3.2 | Adjusted LLM | 11 | 0.779 | 13 | 2.43 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 12 | 0.779 | 11 | 2.40 | Not shortlisted |
| Google Translation LLM | Adjusted NMT | 13 | 0.775 | 12 | 2.41 | Real-time place |
| Translated Lara 2 | Adjusted NMT | 14 | 0.775 | 16 | 2.69 | Real-time place |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 15 | 0.770 | 14 | 2.57 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 16 | 0.766 | 19 | 2.93 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 17 | 0.763 | 15 | 2.60 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 18 | 0.763 | 20 | 3.09 | Not shortlisted |
| Elia Machine Translation System | Off-the-shelf NMT | 19 | 0.762 | 17 | 2.70 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 20 | 0.758 | 18 | 2.73 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 21 | 0.753 | 22 | 3.17 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 22 | 0.740 | 21 | 3.16 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Off-the-shelf LLM | 23 | 0.739 | 23 | 3.18 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 24 | 0.723 | 25 | 3.69 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 25 | 0.722 | 28 | 3.92 | Not shortlisted |
| Google Gemma 4 | Adjusted LLM | 26 | 0.721 | 29 | 4.21 | Not shortlisted |
| MiniMax M2.5 | Off-the-shelf LLM | 27 | 0.718 | 24 | 3.55 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 28 | 0.707 | 27 | 3.82 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 29 | 0.703 | 26 | 3.72 | Not shortlisted |
| Cohere Tiny Aya | Adjusted LLM | 30 | 0.643 | 30 | 5.63 | Not shortlisted |
| AI21 Jamba Large 1.7 | Off-the-shelf LLM | 31 | 0.605 | 31 | 6.10 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 32 | 0.545 | 32 | 8.31 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.794 | — | 2.24 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.790 | — | 2.16 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.765 | — | 2.56 | 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 32 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.
Both metrics rank Google Gemini 3.1 Pro Preview first, yet by Intento LQA the one-step solution on OpenAI GPT-5.6 Sol has fewer weighted errors, 21.8 against 26.4 per 1,000 words. MetricX ranks DeepL Next Gen fourth, and the solution on DeepL Next Gen has the most weighted errors of those with a conclusive Intento LQA result, 127.2 per 1,000 words. Anthropic Claude Sonnet 5 and z.ai GLM-5 made the list through COMET-Kiwi alone. Nearest the cut are DeepSeek V3.2, 11th on COMET-Kiwi, and Anthropic Claude Sonnet 4.6, 11th on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.6 Sol, three-step | $574 |
| Standard | OpenAI GPT-5.6 Sol, three-step | $574 |
| Advanced | OpenAI GPT-5.6 Sol, three-step | $574 |
| Model | US dollars per million characters | Class |
|---|---|---|
| z.ai GLM-5 | $2.01 | LLMs |
| OpenAI GPT-4.1 | $3.11 | LLMs |
| Google Gemini 3.5 Flash | $3.64 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.87 | LLMs |
| Anthropic Claude Sonnet 5 | $6.05 | LLMs |
| OpenAI GPT-5.6 Sol | $7.92 | LLMs |
| OpenAI GPT-5.5 | $11.47 | LLMs |
| Anthropic Claude Opus 4.8 | $15.2 | LLMs |
| Google Translation LLM | $21 | LLM-based MT engines |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $30.3 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Fable 5 | 0.5 | LLMs |
| Anthropic Claude Sonnet 5 | 0.5 | LLMs |
| Anthropic Claude Opus 4.8 | 0.498 | LLMs |
| Mistral Medium 3.1 | 0.415 | LLMs |
| Mistral Large 3 | 0.41 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.4 | LLMs |
| AI21 Jamba Large 1.7 | 0.387 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.379 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.363 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 0.357 | LLMs |
| Cohere Command A Translate | 0.356 | LLMs |
| z.ai GLM-5 | 0.353 | LLMs |
| DeepSeek V3.2 | 0.344 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.338 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.337 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.335 | LLMs |
| Google Gemini 3.5 Flash | 0.334 | LLMs |
| Google Gemma 4 | 0.332 | Open-weight LLMs |
| OpenAI GPT-4.1 | 0.316 | LLMs |
| OpenAI GPT-5.6 Terra | 0.316 | LLMs |
| OpenAI GPT-5.5 | 0.315 | LLMs |
| OpenAI GPT-5.6 Sol | 0.315 | LLMs |
| OpenAI GPT-5.6 Luna | 0.314 | LLMs |
| MiniMax M2.5 | 0.307 | LLMs |
| Cohere Tiny Aya | 0.304 | LLMs |
On OpenAI GPT-5.6 Sol, the one-step solution costs $7.92 per million characters, the two-step one (translation + post-edit) $20.04 and the three-step one (translation + LQA + repair) $574. The tokenizer of Anthropic Claude Sonnet 5 produces 0.5 tokens per character of the translation, against 0.315 for that of OpenAI GPT-5.6 Sol, and yet the one-step solution on Anthropic Claude Sonnet 5 costs less, $6.05 per million characters. Meeting the Advanced requirements costs $574 per million characters, as meeting the Basic ones does.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | OpenAI GPT-5.6 Sol, three-step | 209 |
| Standard | OpenAI GPT-5.6 Sol, three-step | 209 |
| Advanced | OpenAI GPT-5.6 Sol, three-step | 209 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Google Cloud Advanced Translation API | 0.1 | MT engines |
| Microsoft Translator | 0.2 | MT engines |
| Google Translation LLM | 0.5 | LLM-based MT engines |
| DeepL Next Gen | 0.8 | LLM-based MT engines |
| Translated Lara 2 | 1.6 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 2.2 | Open-weight LLMs |
| Elia Machine Translation System | 3.6 | MT engines |
| Yandex Cloud Translate v2 | 3.6 | MT engines |
| Cohere Tiny Aya | 16 | LLMs |
| OpenAI GPT-5.6 Luna | 17 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 23 | LLMs |
| Alibaba Machine Translation API | 28 | MT engines |
| OpenAI GPT-4.1 | 34 | LLMs |
| Mistral Medium 3.1 | 38 | LLMs |
| Swiss AI Apertus-8B-2509 | 40 | Open-weight LLMs |
| Anthropic Claude Haiku 4.5 | 45 | LLMs |
| OpenAI GPT-5.6 Terra | 47 | LLMs |
| MiniMax M2.5 | 51 | LLMs |
| AI21 Jamba Large 1.7 | 58 | LLMs |
| Google Gemini 3.5 Flash | 60 | LLMs |
| Mistral Large 3 | 60 | LLMs |
| Cohere Command A Translate | 62 | LLMs |
| xAI Grok 4.20 (reasoning) | 65 | LLMs |
| DeepSeek V3.2 | 68 | LLMs |
| Anthropic Claude Sonnet 5 | 80 | LLMs |
| Google Gemini 3.1 Pro Preview | 86 | LLMs |
| Anthropic Claude Sonnet 4.6 | 86 | LLMs |
| OpenAI GPT-5.6 Sol | 87 | LLMs |
| Anthropic Claude Fable 5 | 88 | LLMs |
| OpenAI GPT-5.5 | 89 | LLMs |
| Anthropic Claude Opus 4.8 | 105 | LLMs |
| Alibaba Qwen3-235B-A22B | 117 | Open-weight LLMs |
| Google Gemma 4 | 167 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 182 | LLMs |
| z.ai GLM-5 | 265 | 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 Basque.
The real-time models for Basque are three LLM-based MT engines, two other MT engines and the open-weight OpenAI gpt-oss-120b. The three LLM-based MT engines are on the shortlist. Translations from DeepL Next Gen and Google Translation LLM have 94.1% and 95.1% of the tags right. The one-step LLM solutions on the shortlist take 34 to 265 seconds per 1,000 words. At every quality level, the fastest solution that meets it is the three-step solution (translation + LQA + repair) on OpenAI GPT-5.6 Sol, at 209 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% | ✓ | $574 | 209 |
| Google Gemini 3.1 Pro Preview | three-step | 100% | ✓ | $577 | 273 |
For Basque we check the formal register, with the form of address zu (zuka), and gender neutrality: no gender absent from the English source and no default gender. Basque has no grammatical gender for persons in the formal register, so formal translations are gender-neutral by default, gendered terms aside. Every shortlisted solution meets tone of voice. Both three-step solutions (translation + LQA + repair) that meet every Advanced requirement have 100% of the sentences about people gender-neutral without the gender-neutrality agent.
For the 12 shortlisted solutions in Basque, the median share of sentences on which the three LQA judges agree is 91.2%. The quality-estimation pair of automatic metrics has an agreement of 0.67 against a floor of 0.53, and the similarity pair 0.59 against a floor of 0.46. The low ends of their ranges, 0.66 and 0.58, also stay above the floors. So the ranking of the shortlist and the Intento LQA error counts on this page carry weight in Basque.
Leave feedback