The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Italian 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 (Lei) 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 Italian, 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. 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.
The four two-step solutions (translation + post-edit) that meet every Standard requirement in Italian have 14.3 to 20.0 total errors per 1,000 words. The one on OpenAI GPT-5.5 has the fewest, at $31.25 per million characters and 89 seconds per 1,000 words, and Optimize for selects it for quality. The one on OpenAI GPT-5.6 Luna has 20.0 total errors per 1,000 words, at $0.41 per million characters and 12 seconds per 1,000 words, and Optimize for selects it for cost and latency.
Under the Standard requirements, only the two-step solutions on OpenAI GPT-5.6 Luna and Google Gemini 3.1 Pro Preview cost less than $31.25 per million characters, at $0.41 and $4.11. Of these two, the one on OpenAI GPT-5.6 Luna takes 12 seconds per 1,000 words, against 82 for the one on Google Gemini 3.1 Pro Preview.
For the Advanced requirements, the gender-neutrality agent works with the two-step solutions (translation + post-edit) on OpenAI GPT-5.5 and Anthropic Claude Opus 4.8. On OpenAI GPT-5.5 the agent takes the cost from $31.25 to $37.56 per million characters and the time from 89 to 117 seconds per 1,000 words. Total errors fall from 14.3 to 12.2 per 1,000 words. On Anthropic Claude Opus 4.8 the cost goes from $31.75 to $32.74 per million characters and total errors from 17.0 to 12.2 per 1,000 words.
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Luna
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Luna
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
two-step
OpenAI GPT-5.6 Luna
two-step
OpenAI GPT-5.6 Luna
two-step
OpenAI GPT-5.5
two-step with the gender-neutrality agent
OpenAI GPT-5.5
two-step with the gender-neutrality agent
OpenAI GPT-5.5
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 | ★×5 | one-step | OpenAI GPT-5.6 Luna | 20.7 | $0.41 | 12 |
| Standard | Tier 2: two or more solutions, none simple | ☆×4 | two-step | OpenAI GPT-5.6 Luna | 20.0 | $0.41 | 12 |
| Advanced | Tier 3: one solution | ☆ | with the gender-neutrality agent | OpenAI GPT-5.5 | 12.2 | $37.56 | 117 |
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.
Five one-step solutions, all on LLMs, meet every Basic requirement in Italian. Optimize for selects the one on OpenAI GPT-5.5 for quality, with 14.0 total errors per 1,000 words, and the one on OpenAI GPT-5.6 Luna for cost and latency. The Standard requirements take two steps (translation + post-edit) and remove the solution on Cohere Command A Translate. In two steps, translations from it have 97.9% of the sentences in the formal register and 98.3% of the glossary terms right.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.13 | 28.7 | Misses a Standard requirement |
| OpenAI GPT-5.6 Luna | 0.41 | 17.5 | Misses a Standard requirement |
| Alibaba Qwen3-235B-A22B | 1.09 | 44.5 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.11 | 15.5 | Misses a Standard requirement |
| Anthropic Claude Sonnet 4.6 | 6.3 | 26.7 | Misses a Standard requirement |
| Widn AI Tower Sugarloaf | 10 | 35.6 | Misses a Standard requirement |
| Microsoft Translator | 10 | 62.6 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 10.26 | 11.2 | Misses a Standard requirement |
| Anthropic Claude Opus 4.8 | 14.52 | 14.5 | Misses a Standard requirement |
| Amazon Translate | 15 | 72.8 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 20.4 | Misses a Standard requirement |
Among the one-step solutions with a cost, the dearest, on DeepL Next Gen at $27.50 per million characters, has 20.4 weighted errors per 1,000 words, against 11.2 for the one on OpenAI GPT-5.5 at $10.26.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 23.5 |
| Tone of voice | 3.1 |
| Gender neutrality | 2.4 |
| Tag handling | 1.3 |
| Terminology | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
Tone of voice is the Standard requirement the fewest one-step solutions meet in Italian. Translations from four LLMs, one of them the open-weight Alibaba Qwen3-235B-A22B, have 97.9% of the sentences in the formal register, and those from Cohere Command A Translate 74.5%. Those from the LLM-based MT engines have 83.0% and 91.5% of the sentences in the formal register, and those from the MT engines 17.0% and 4.3%. In two steps (translation + post-edit), the solutions on OpenAI GPT-5.5 and Anthropic Claude Opus 4.8 meet tone of voice.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Italian | Class |
|---|---|---|
| Amazon Translate | 72.8 | MT engines |
| Microsoft Translator | 62.6 | MT engines |
| Alibaba Qwen3-235B-A22B | 44.5 | Open-weight LLMs |
| Widn AI Tower Sugarloaf | 35.6 | LLM-based MT engines |
| Google Gemma 4 | 28.7 | Open-weight LLMs |
| Anthropic Claude Sonnet 4.6 | 26.7 | LLMs |
| DeepL Next Gen | 20.4 | LLM-based MT engines |
| OpenAI GPT-5.6 Luna | 17.5 | LLMs |
| Cohere Command A Translate | 16.5 | LLMs |
| Google Gemini 3.1 Pro Preview | 15.5 | LLMs |
| Anthropic Claude Opus 4.8 | 14.5 | LLMs |
| OpenAI GPT-5.5 | 11.2 | LLMs |
No translation into Italian: SpeakLeash Bielik-11B-v3.0-Instruct (wrong language), Widn AI Tower Vesuvius (cut off part-way).
The five one-step solutions that meet general translation quality in Italian are all on LLMs. In one step, translations from the six shortlisted LLMs that are not open-weight have 11.2 to 26.7 weighted errors per 1,000 words, and those from the open-weight Google Gemma 4 and Alibaba Qwen3-235B-A22B 28.7 and 44.5. Those from the LLM-based MT engines have 20.4 and 35.6 weighted errors per 1,000 words, and those from the MT engines 62.6 and 72.8. All eight shortlisted LLMs support Italian by our measured criteria.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Anthropic Claude Fable 5 | 0.808 | 10 | 1.70 | 22 | No | Not shortlisted |
| OpenAI GPT-4.1 | 0.808 | 11 | 1.60 | 11 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.805 | 13 | 1.59 | 9 | No | Not shortlisted |
| Google Gemini 3.5 Flash | 0.805 | 14 | 1.70 | 23 | No | Not shortlisted |
| Mistral Large 3 | 0.805 | 15 | 1.72 | 24 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.804 | 16 | 1.72 | 25 | No | Not shortlisted |
| Tilde TildeOpen LLM engine | 0.803 | 17 | 1.63 | 13 | No | Not shortlisted |
| Supertext LLM | 0.802 | 18 | 1.68 | 21 | No | Not shortlisted |
| DeepSeek V3.2 | 0.802 | 19 | 1.81 | 29 | No | Not shortlisted |
| MiniMax M2.5 | 0.802 | 20 | 1.79 | 27 | No | Not shortlisted |
| Anthropic Claude Sonnet 5 | 0.801 | 21 | 1.67 | 19 | No | Not shortlisted |
| Translated Lara 2 | 0.801 | 22 | 1.86 | 33 | No | Not shortlisted |
| Meta Llama 4 Maverick | 0.800 | 24 | 1.80 | 28 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.800 | 26 | 1.64 | 14 | No | Not shortlisted |
| z.ai GLM-5 | 0.800 | 27 | 1.66 | 18 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.798 | 28 | 1.65 | 16 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.796 | 29 | 1.82 | 30 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.796 | 30 | 1.64 | 15 | No | Not shortlisted |
| Google Translation LLM | 0.795 | 31 | 2.10 | 37 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.793 | 32 | 1.84 | 31 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.790 | 33 | 1.90 | 34 | No | Not shortlisted |
| Pangeanic PangeaMT | 0.790 | 34 | 1.75 | 26 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.778 | 35 | 1.86 | 32 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.776 | 36 | 2.40 | 41 | No | Not shortlisted |
| Cohere Tiny Aya | 0.772 | 37 | 2.16 | 38 | No | Not shortlisted |
| Oracle Cloud Language Service | 0.768 | 38 | 2.04 | 35 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.766 | 39 | 2.19 | 39 | No | Not shortlisted |
| Baidu Translate API | 0.765 | 40 | 2.07 | 36 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.754 | 41 | 2.31 | 40 | No | Not shortlisted |
| Tencent Cloud TMT | 0.751 | 42 | 2.52 | 42 | No | Not shortlisted |
| AppTek NMT | 0.693 | 43 | 3.54 | 43 | No | Not shortlisted |
| Widn AI Tower Anthill | 0.458 | 44 | 11.14 | 44 | No | Not shortlisted |
| Cohere Command A Translate | 0.817 | 1 | 1.46 | 1 | Yes | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | 0.816 | 2 | 1.60 | 10 | Yes | Top 8 by COMET-Kiwi |
| DeepL Next Gen | 0.815 | 3 | 1.48 | 3 | Yes | Top 8 by both metrics |
| Widn AI Tower Sugarloaf | 0.811 | 4 | 1.55 | 7 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Luna | 0.810 | 5 | 1.61 | 12 | Yes | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.5 | 0.809 | 6 | 1.68 | 20 | Yes | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | 0.809 | 7 | 1.65 | 17 | Yes | Top 8 by COMET-Kiwi |
| Amazon Translate | 0.808 | 8 | 1.49 | 4 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.808 | 9 | 1.48 | 2 | Yes | Top 8 by MetricX |
| Anthropic Claude Sonnet 4.6 | 0.807 | 12 | 1.54 | 5 | Yes | Top 8 by MetricX |
| Microsoft Translator | 0.800 | 23 | 1.57 | 8 | Yes | Top 8 by MetricX |
| Alibaba Qwen3-235B-A22B | 0.800 | 25 | 1.54 | 6 | Yes | Top 8 by MetricX |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| Cohere Command A Translate | Adjusted LLM | 1 | 0.817 | 1 | 1.46 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 2 | 0.816 | 10 | 1.60 | Top 8 by COMET-Kiwi |
| DeepL Next Gen | Adjusted NMT | 3 | 0.815 | 3 | 1.48 | Top 8 by both metrics |
| Widn AI Tower Sugarloaf | Adjusted NMT | 4 | 0.811 | 7 | 1.55 | Top 8 by both metrics |
| OpenAI GPT-5.6 Luna | Adjusted LLM | 5 | 0.810 | 12 | 1.61 | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.5 | Adjusted LLM | 6 | 0.809 | 20 | 1.68 | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 7 | 0.809 | 17 | 1.65 | Top 8 by COMET-Kiwi |
| Amazon Translate | Off-the-shelf NMT | 8 | 0.808 | 4 | 1.49 | Top 8 by both metrics |
| Google Gemma 4 | Adjusted LLM | 9 | 0.808 | 2 | 1.48 | Top 8 by MetricX |
| Anthropic Claude Fable 5 | Adjusted LLM | 10 | 0.808 | 22 | 1.70 | Not shortlisted |
| OpenAI GPT-4.1 | Adjusted LLM | 11 | 0.808 | 11 | 1.60 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 12 | 0.807 | 5 | 1.54 | Top 8 by MetricX |
| ModernMT Enterprise Edition | Adjusted NMT | 13 | 0.805 | 9 | 1.59 | Not shortlisted |
| Google Gemini 3.5 Flash | Adjusted LLM | 14 | 0.805 | 23 | 1.70 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 15 | 0.805 | 24 | 1.72 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 16 | 0.804 | 25 | 1.72 | Not shortlisted |
| Tilde TildeOpen LLM engine | Adjusted NMT | 17 | 0.803 | 13 | 1.63 | Not shortlisted |
| Supertext LLM | Off-the-shelf NMT | 18 | 0.802 | 21 | 1.68 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 19 | 0.802 | 29 | 1.81 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 20 | 0.802 | 27 | 1.79 | Not shortlisted |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 21 | 0.801 | 19 | 1.67 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 22 | 0.801 | 33 | 1.86 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 23 | 0.800 | 8 | 1.57 | Top 8 by MetricX |
| Meta Llama 4 Maverick | Adjusted LLM | 24 | 0.800 | 28 | 1.80 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 25 | 0.800 | 6 | 1.54 | Top 8 by MetricX |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 26 | 0.800 | 14 | 1.64 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 27 | 0.800 | 18 | 1.66 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 28 | 0.798 | 16 | 1.65 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 29 | 0.796 | 30 | 1.82 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 30 | 0.796 | 15 | 1.64 | Not shortlisted |
| Google Translation LLM | Off-the-shelf NMT | 31 | 0.795 | 37 | 2.10 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 32 | 0.793 | 31 | 1.84 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 33 | 0.790 | 34 | 1.90 | Not shortlisted |
| Pangeanic PangeaMT | Off-the-shelf NMT | 34 | 0.790 | 26 | 1.75 | Not shortlisted |
| SYSTRAN PNMT API | Adjusted NMT | 35 | 0.778 | 32 | 1.86 | Not shortlisted |
| Google Cloud Advanced Translation API | Adjusted NMT | 36 | 0.776 | 41 | 2.40 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 37 | 0.772 | 38 | 2.16 | Not shortlisted |
| Oracle Cloud Language Service | Off-the-shelf NMT | 38 | 0.768 | 35 | 2.04 | Not shortlisted |
| Yandex Cloud Translate v2 | Adjusted NMT | 39 | 0.766 | 39 | 2.19 | Not shortlisted |
| Baidu Translate API | Off-the-shelf NMT | 40 | 0.765 | 36 | 2.07 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 41 | 0.754 | 40 | 2.31 | Not shortlisted |
| Tencent Cloud TMT | Off-the-shelf NMT | 42 | 0.751 | 42 | 2.52 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 43 | 0.693 | 43 | 3.54 | Not shortlisted |
| Widn AI Tower Anthill | Adjusted NMT | 44 | 0.458 | 44 | 11.14 | Not shortlisted |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.807 | — | 1.57 | Same model family as OpenAI GPT-5.6 Luna |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.804 | — | 1.64 | Same model family as OpenAI GPT-5.6 Luna |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.801 | — | 1.67 | 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 44 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.
Four solutions are on the Italian shortlist through MetricX alone, on Google Gemma 4, Anthropic Claude Sonnet 4.6, Alibaba Qwen3-235B-A22B and Microsoft Translator. None of them meet general translation quality, with 26.7 to 62.6 weighted errors per 1,000 words. Of the four in the top 8 of both metrics, only the one on Cohere Command A Translate meets it. Nearest the cut in the tail are Anthropic Claude Fable 5, 10th on COMET-Kiwi, and ModernMT Enterprise Edition, 9th on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.6 Luna, one-step LLM adjusted for Standard requirements | $0.41 |
| Standard | OpenAI GPT-5.6 Luna, two-step | $0.41 |
| Advanced | OpenAI GPT-5.5, two-step with the gender-neutrality agent | $37.56 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.13 | Open-weight LLMs |
| OpenAI GPT-5.6 Luna | $0.41 | LLMs |
| Alibaba Qwen3-235B-A22B | $1.09 | Open-weight LLMs |
| Google Gemini 3.1 Pro Preview | $4.11 | LLMs |
| Anthropic Claude Sonnet 4.6 | $6.3 | LLMs |
| Widn AI Tower Sugarloaf | $10 | LLM-based MT engines |
| Microsoft Translator | $10 | MT engines |
| OpenAI GPT-5.5 | $10.26 | LLMs |
| Anthropic Claude Opus 4.8 | $14.52 | LLMs |
| Amazon Translate | $15 | MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Fable 5 | 0.429 | LLMs |
| Anthropic Claude Opus 4.8 | 0.429 | LLMs |
| Anthropic Claude Sonnet 5 | 0.429 | LLMs |
| Mistral Medium 3.1 | 0.343 | LLMs |
| Mistral Large 3 | 0.333 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.317 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.316 | LLMs |
| MoonshotAI Kimi K2.5 | 0.304 | LLMs |
| AI21 Jamba Large 1.7 | 0.3 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.294 | Open-weight LLMs |
| Widn AI Tower Sugarloaf | 0.293 | LLM-based MT engines |
| DeepSeek V3.2 | 0.283 | LLMs |
| z.ai GLM-5 | 0.274 | LLMs |
| OpenAI GPT-4.1 | 0.264 | LLMs |
| OpenAI GPT-5.5 | 0.264 | LLMs |
| Cohere Command A Translate | 0.263 | LLMs |
| OpenAI GPT-5.6 Sol | 0.262 | LLMs |
| OpenAI GPT-5.6 Terra | 0.262 | LLMs |
| OpenAI GPT-5.6 Luna | 0.261 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.259 | LLMs |
| Google Gemma 4 | 0.258 | Open-weight LLMs |
| Google Gemini 3.5 Flash | 0.257 | LLMs |
| MiniMax M2.5 | 0.251 | LLMs |
| Cohere Tiny Aya | 0.244 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.241 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.239 | LLMs |
The tokenizers of OpenAI GPT-5.5, OpenAI GPT-5.6 Luna and Google Gemini 3.1 Pro Preview produce 0.264, 0.261 and 0.259 tokens per character of the Italian translation. Yet the one-step solutions on them cost $10.26, $0.41 and $4.11 per million characters. On OpenAI GPT-5.5 the cost follows the steps: $10.26 per million characters in one step, $31.25 in two steps (translation + post-edit) and $37.56 with the gender-neutrality agent.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | OpenAI GPT-5.6 Luna, one-step LLM adjusted for Standard requirements | 12 |
| Standard | OpenAI GPT-5.6 Luna, two-step | 12 |
| Advanced | OpenAI GPT-5.5, two-step with the gender-neutrality agent | 117 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.2 | MT engines |
| Google Cloud Advanced Translation API | 0.2 | MT engines |
| Google Translation LLM | 0.5 | LLM-based MT engines |
| Pangeanic PangeaMT | 0.5 | MT engines |
| DeepL Next Gen | 1.1 | LLM-based MT engines |
| Amazon Translate | 1.2 | MT engines |
| Translated Lara 2 | 1.2 | LLM-based MT engines |
| ModernMT Enterprise Edition | 1.3 | MT engines |
| SYSTRAN PNMT API | 1.3 | MT engines |
| OpenAI gpt-oss-120b | 1.4 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 2.6 | MT engines |
| Oracle Cloud Language Service | 3.3 | MT engines |
| Widn AI Tower Anthill | 6.9 | LLM-based MT engines |
| Tilde TildeOpen LLM engine | 8 | LLM-based MT engines |
| OpenAI GPT-5.6 Luna | 12 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 12 | LLMs |
| Supertext LLM | 14 | LLM-based MT engines |
| Baidu Translate API | 15 | MT engines |
| Cohere Tiny Aya | 18 | LLMs |
| OpenAI GPT-4.1 | 19 | LLMs |
| Alibaba Machine Translation API | 19 | MT engines |
| OpenAI GPT-5.6 Terra | 23 | LLMs |
| Anthropic Claude Haiku 4.5 | 25 | LLMs |
| Cohere Command A Translate | 28 | LLMs |
| Meta Llama 4 Maverick | 30 | Open-weight LLMs |
| Mistral Medium 3.1 | 30 | LLMs |
| DeepSeek V3.2 | 34 | LLMs |
| MiniMax M2.5 | 35 | LLMs |
| AppTek NMT | 35 | MT engines |
| Anthropic Claude Sonnet 5 | 37 | LLMs |
| OpenAI GPT-5.6 Sol | 38 | LLMs |
| Widn AI Tower Sugarloaf | 40 | LLM-based MT engines |
| OpenAI GPT-5.5 | 40 | LLMs |
| xAI Grok 4.20 (reasoning) | 41 | LLMs |
| AI21 Jamba Large 1.7 | 41 | LLMs |
| Mistral Large 3 | 44 | LLMs |
| Anthropic Claude Fable 5 | 45 | LLMs |
| Anthropic Claude Opus 4.8 | 50 | LLMs |
| Anthropic Claude Sonnet 4.6 | 51 | LLMs |
| Tencent Cloud TMT | 53 | MT engines |
| Google Gemini 3.5 Flash | 57 | LLMs |
| Swiss AI Apertus-8B-2509 | 59 | Open-weight LLMs |
| Google Gemini 3.1 Pro Preview | 82 | LLMs |
| Alibaba Qwen3-235B-A22B | 101 | Open-weight LLMs |
| z.ai GLM-5 | 174 | LLMs |
| MoonshotAI Kimi K2.5 | 175 | LLMs |
| Google Gemma 4 | 200 | Open-weight 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 Italian.
The 10 real-time models in Italian are six MT engines, three LLM-based MT engines and OpenAI gpt-oss-120b. Translations from the shortlisted Microsoft Translator, DeepL Next Gen and Amazon Translate have 4.3%, 83.0% and 17.0% of the sentences in the formal register. The fastest solutions are on OpenAI GPT-5.6 Luna, in one step under the Basic requirements and in two steps (translation + post-edit) under the Standard ones, at 12 seconds per 1,000 words. Under the Advanced ones the only solution takes 117 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-5.5 | two-step, with the gender-neutrality agent | 100% | ✓ | $37.56 | 117 |
| Anthropic Claude Opus 4.8 | two-step, with the gender-neutrality agent | 99.1% | ✗ | $32.74 | 93 |
For Italian we check the formal register, with Lei as the form of address. Gender-neutral wording here means neutral nouns and rephrasing, doubling only when unavoidable, no typographic forms such as ə or the asterisk and no default gender. The gender-neutrality agent works with both of the two best solutions. With it, the two-step solution (translation + post-edit) on Anthropic Claude Opus 4.8 meets every Advanced requirement but gender neutrality. There we require every sentence about people to be gender-neutral.
The three LQA judges agree on a median of 96.0% of the sentences across the 12 shortlisted solutions in Italian. The error counts on this page rest on agreement that close. With both pairs of automatic metrics in agreement as well, the scores behind the shortlist carry weight too.
Leave feedback