The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Arabic (MSA) 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 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 Arabic (MSA), 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. 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 one-step solutions meet every Standard requirement in Arabic (MSA), and neither leads on errors, cost and time at once. The one on OpenAI GPT-5.5 has 43.9 total errors per 1,000 words against 44.5 for the one on Anthropic Claude Fable 5, and costs $10.27 per million characters against $32.52. The one on Anthropic Claude Fable 5 takes 58 seconds per 1,000 words against 67. Optimize for selects OpenAI GPT-5.5 for quality and cost, and Anthropic Claude Fable 5 for latency.
None of the solutions that meet every Standard requirement in Arabic (MSA) cost less than $10.27 per million characters or take less than 57 seconds per 1,000 words. A cost cap between $10.28 and $32.51 per million characters leaves only the solution on OpenAI GPT-5.5, and a time limit between 58 and 67 seconds per 1,000 words only the one on Anthropic Claude Fable 5. Both are one-step solutions, so a limit of one step keeps both.
With the gender-neutrality agent, both solutions that meet every Standard requirement become three-step solutions (translation + LQA + repair), still short of the Advanced requirements. With the two added steps and the agent together, their total errors fall from 43.9 to 10.4 per 1,000 words on OpenAI GPT-5.5 and from 44.5 to 8.9 on Anthropic Claude Fable 5. Without the agent, 27.0% and 29.5% of the sentences about people in their one-step translations are gender-neutral.
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
three-step with the gender-neutrality agent
Misses gender neutrality.
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 | OpenAI GPT-5.5 | 43.9 | $10.27 | 67 |
| Standard | Tier 1: simple solution + backup | ★★ | one-step | OpenAI GPT-5.5 | 43.9 | $10.27 | 67 |
| Advanced | Tier 4: no solution | No solution | — | — | — |
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.
Two one-step solutions, on OpenAI GPT-5.5 and Anthropic Claude Fable 5, meet every Basic and every Standard requirement in Arabic (MSA), so one step is enough. In both cases, Optimize for picks OpenAI GPT-5.5 for quality and cost and Anthropic Claude Fable 5 for latency. With the gender-neutrality agent, the three-step solution (translation + LQA + repair) on OpenAI GPT-5.5 misses terminology as well as gender neutrality, with 98.8% of the glossary terms right. The one on Anthropic Claude Fable 5 misses gender neutrality alone.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.13 | 43.2 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 2.88 | 35.6 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.28 | 33.6 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.38 | 25.4 | Misses a Standard requirement |
| Anthropic Claude Sonnet 5 | 6.52 | 29.3 | Misses a Standard requirement |
| Anthropic Claude Sonnet 4.6 | 9.04 | 45.8 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 10.27 | 16.2 | Meets every Standard requirement |
| Google Translation LLM | 19.81 | 82.1 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 30.7 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 32.52 | 17.8 | Meets every Standard requirement |
| ALL-Lab | 100 | 29.7 | Misses a Standard requirement |
None of the one-step solutions in Arabic (MSA) that cost more than the one on OpenAI GPT-5.5 have fewer weighted errors. It costs $10.27 per million characters and has 16.2 weighted errors per 1,000 words. Those on ALL-Lab, Anthropic Claude Fable 5, DeepL Next Gen and Google Translation LLM cost $100, $32.52, $27.50 and $19.81 per million characters and have 29.7, 17.8, 30.7 and 82.1 weighted errors per 1,000 words. The six that cost less, from $0.13 to $9.04 per million characters, have 25.4 to 45.8 weighted errors per 1,000 words.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 30.7 |
| Gender neutrality | 26.7 |
| Tag handling | 0.7 |
| Tone of voice | 0.2 |
| Terminology | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
Translations from Google Gemini 3.1 Pro Preview, Anthropic Claude Sonnet 5, Google Gemini 3.5 Flash and Google Gemma 4 have 99.6% of the tags right, under the 100% that tag handling requires. The one-step solution closest to the bar of general translation quality, on Google Gemini 3.1 Pro Preview, has 25.4 weighted errors per 1,000 words and no critical error. In translations from the off-the-shelf solutions, ALL-Lab and Google Translation LLM, 75.9% and 78.5% of the glossary terms are right, and 97.6% to 100% in the adjusted ones.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Arabic (MSA) | Class |
|---|---|---|
| Google Translation LLM | 82.1 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 72.2 | Open-weight LLMs |
| Anthropic Claude Sonnet 4.6 | 45.8 | LLMs |
| Google Gemma 4 | 43.2 | Open-weight LLMs |
| OpenAI GPT-4.1 | 35.6 | LLMs |
| Google Gemini 3.5 Flash | 33.6 | LLMs |
| DeepL Next Gen | 30.7 | LLM-based MT engines |
| ALL-Lab ALL-Lab | 29.7 | MT engines |
| Anthropic Claude Sonnet 5 | 29.3 | LLMs |
| Cohere Command A Translate | 27.4 | LLMs |
| Google Gemini 3.1 Pro Preview | 25.4 | LLMs |
| Anthropic Claude Fable 5 | 17.8 | LLMs |
| OpenAI GPT-5.5 | 16.2 | LLMs |
No translation into Arabic (MSA): Widn AI Tower Anthill (far too short, wrong language, wrong script), Widn AI Tower Vesuvius (far too short, wrong language, wrong script).
The one-step solutions on the eight shortlisted LLMs that are not open-weight, among them the two that meet general translation quality, have 16.2 to 45.8 weighted errors per 1,000 words in Arabic (MSA). Translations from ALL-Lab, the one shortlisted MT engine, have 29.7 weighted errors per 1,000 words, and those from the LLM-based MT engines DeepL Next Gen and Google Translation LLM 30.7 and 82.1. The shortlisted LLM outside our language-support criteria is the open-weight OpenAI gpt-oss-120b, at 72.2 weighted errors per 1,000 words.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| MoonshotAI Kimi K2.5 | 0.783 | 10 | 1.35 | 12 | No | Not shortlisted |
| Anthropic Claude Opus 4.8 | 0.783 | 12 | 1.36 | 13 | No | Not shortlisted |
| OpenAI GPT-5.6 Terra | 0.782 | 13 | 1.37 | 15 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.782 | 14 | 1.40 | 18 | No | Not shortlisted |
| z.ai GLM-5 | 0.781 | 15 | 1.30 | 11 | No | Not shortlisted |
| DeepSeek V3.2 | 0.779 | 16 | 1.36 | 14 | No | Not shortlisted |
| Mistral Large 3 | 0.779 | 17 | 1.39 | 16 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.778 | 18 | 1.44 | 21 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.775 | 20 | 1.39 | 17 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.773 | 22 | 1.42 | 20 | No | Not shortlisted |
| Translated Lara 2 | 0.770 | 24 | 1.52 | 24 | No | Not shortlisted |
| Meta Llama 4 Maverick | 0.770 | 25 | 1.69 | 30 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.770 | 26 | 1.54 | 26 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.768 | 27 | 1.56 | 27 | No | Not shortlisted |
| Microsoft Translator | 0.767 | 28 | 1.51 | 23 | No | Not shortlisted |
| Amazon Translate | 0.766 | 29 | 1.57 | 28 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.765 | 30 | 1.52 | 25 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.764 | 31 | 1.64 | 29 | No | Not shortlisted |
| Tarjama NMT | 0.760 | 32 | 1.88 | 33 | No | Not shortlisted |
| Pangeanic PangeaMT | 0.753 | 33 | 1.75 | 32 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.739 | 34 | 2.44 | 35 | No | Not shortlisted |
| Oracle Cloud Language Service | 0.734 | 35 | 2.01 | 34 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.718 | 36 | 2.45 | 36 | No | Not shortlisted |
| MiniMax M2.5 | 0.702 | 37 | 2.96 | 38 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.700 | 38 | 2.82 | 37 | No | Not shortlisted |
| iFlyTek MT | 0.673 | 39 | 3.53 | 39 | No | Not shortlisted |
| Cohere Tiny Aya | 0.647 | 40 | 4.52 | 40 | No | Not shortlisted |
| AppTek NMT | 0.590 | 41 | 6.13 | 41 | No | Not shortlisted |
| DeepL Next Gen | 0.793 | 1 | 1.14 | 1 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.791 | 2 | 1.30 | 10 | Yes | Top 8 by COMET-Kiwi |
| Cohere Command A Translate | 0.788 | 3 | 1.26 | 6 | Yes | Top 8 by both metrics |
| OpenAI GPT-4.1 | 0.786 | 4 | 1.23 | 3 | Yes | Top 8 by both metrics |
| ALL-Lab | 0.786 | 5 | 1.40 | 19 | Yes | Top 8 by COMET-Kiwi |
| Google Gemma 4 | 0.786 | 6 | 1.24 | 4 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.785 | 7 | 1.23 | 2 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | 0.785 | 8 | 1.30 | 9 | Yes | Top 8 by COMET-Kiwi |
| Google Gemini 3.5 Flash | 0.783 | 9 | 1.24 | 5 | Yes | Top 8 by MetricX |
| OpenAI GPT-5.5 | 0.783 | 11 | 1.29 | 8 | Yes | Top 8 by MetricX |
| Google Translation LLM | 0.775 | 19 | 1.69 | 31 | Yes | Real-time place |
| Anthropic Claude Sonnet 4.6 | 0.774 | 21 | 1.28 | 7 | Yes | Top 8 by MetricX |
| OpenAI gpt-oss-120b | 0.773 | 23 | 1.44 | 22 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| DeepL Next Gen | Adjusted NMT | 1 | 0.793 | 1 | 1.14 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 2 | 0.791 | 10 | 1.30 | Top 8 by COMET-Kiwi |
| Cohere Command A Translate | Adjusted LLM | 3 | 0.788 | 6 | 1.26 | Top 8 by both metrics |
| OpenAI GPT-4.1 | Adjusted LLM | 4 | 0.786 | 3 | 1.23 | Top 8 by both metrics |
| ALL-Lab | Off-the-shelf NMT | 5 | 0.786 | 19 | 1.40 | Top 8 by COMET-Kiwi |
| Google Gemma 4 | Adjusted LLM | 6 | 0.786 | 4 | 1.24 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 7 | 0.785 | 2 | 1.23 | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 8 | 0.785 | 9 | 1.30 | Top 8 by COMET-Kiwi |
| Google Gemini 3.5 Flash | Adjusted LLM | 9 | 0.783 | 5 | 1.24 | Top 8 by MetricX |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 10 | 0.783 | 12 | 1.35 | Not shortlisted |
| OpenAI GPT-5.5 | Adjusted LLM | 11 | 0.783 | 8 | 1.29 | Top 8 by MetricX |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 12 | 0.783 | 13 | 1.36 | Not shortlisted |
| OpenAI GPT-5.6 Terra | Adjusted LLM | 13 | 0.782 | 15 | 1.37 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 14 | 0.782 | 18 | 1.40 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 15 | 0.781 | 11 | 1.30 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 16 | 0.779 | 14 | 1.36 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 17 | 0.779 | 16 | 1.39 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 18 | 0.778 | 21 | 1.44 | Not shortlisted |
| Google Translation LLM | Off-the-shelf NMT | 19 | 0.775 | 31 | 1.69 | Real-time place |
| AI21 Jamba Large 1.7 | Adjusted LLM | 20 | 0.775 | 17 | 1.39 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 21 | 0.774 | 7 | 1.28 | Top 8 by MetricX |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 22 | 0.773 | 20 | 1.42 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 23 | 0.773 | 22 | 1.44 | Real-time place |
| Translated Lara 2 | Adjusted NMT | 24 | 0.770 | 24 | 1.52 | Not shortlisted |
| Meta Llama 4 Maverick | Adjusted LLM | 25 | 0.770 | 30 | 1.69 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 26 | 0.770 | 26 | 1.54 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 27 | 0.768 | 27 | 1.56 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 28 | 0.767 | 23 | 1.51 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 29 | 0.766 | 28 | 1.57 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 30 | 0.765 | 25 | 1.52 | Not shortlisted |
| SYSTRAN PNMT API | Adjusted NMT | 31 | 0.764 | 29 | 1.64 | Not shortlisted |
| Tarjama NMT | Off-the-shelf NMT | 32 | 0.760 | 33 | 1.88 | Not shortlisted |
| Pangeanic PangeaMT | Off-the-shelf NMT | 33 | 0.753 | 32 | 1.75 | Not shortlisted |
| Google Cloud Advanced Translation API | Adjusted NMT | 34 | 0.739 | 35 | 2.44 | Not shortlisted |
| Oracle Cloud Language Service | Off-the-shelf NMT | 35 | 0.734 | 34 | 2.01 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 36 | 0.718 | 36 | 2.45 | Not shortlisted |
| MiniMax M2.5 | Off-the-shelf LLM | 37 | 0.702 | 38 | 2.96 | Not shortlisted |
| Yandex Cloud Translate v2 | Adjusted NMT | 38 | 0.700 | 37 | 2.82 | Not shortlisted |
| iFlyTek MT | Off-the-shelf NMT | 39 | 0.673 | 39 | 3.53 | Not shortlisted |
| Cohere Tiny Aya | Adjusted LLM | 40 | 0.647 | 40 | 4.52 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 41 | 0.590 | 41 | 6.13 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.782 | — | 1.48 | Same model family as OpenAI GPT-5.6 Terra |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.778 | — | 1.36 | Same model family as OpenAI GPT-5.6 Terra |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.781 | — | 1.42 | 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 41 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.
Each of the two solutions that meet every Standard requirement in Arabic (MSA) is on the shortlist through one metric. Anthropic Claude Fable 5 is at rank 2 on COMET-Kiwi and rank 10 on MetricX, and OpenAI GPT-5.5 at rank 8 on MetricX and rank 11 on COMET-Kiwi. The one-step solution on Google Gemini 3.1 Pro Preview, at rank 2 on MetricX, has 25.4 weighted errors per 1,000 words. Nearest the cut in the tail are MoonshotAI Kimi K2.5, at rank 10 on COMET-Kiwi, and z.ai GLM-5, at rank 11 on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.5, one-step LLM adjusted for Standard requirements | $10.27 |
| Standard | OpenAI GPT-5.5, one-step LLM adjusted for Standard requirements | $10.27 |
| Advanced | No solution | — |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.13 | Open-weight LLMs |
| OpenAI GPT-4.1 | $2.88 | LLMs |
| Google Gemini 3.5 Flash | $3.28 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.38 | LLMs |
| Anthropic Claude Sonnet 5 | $6.52 | LLMs |
| Anthropic Claude Sonnet 4.6 | $9.04 | LLMs |
| OpenAI GPT-5.5 | $10.27 | LLMs |
| Google Translation LLM | $19.81 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $32.52 | LLMs |
| ALL-Lab | $100 | MT engines |
| Model | Tokens per character | Class |
|---|---|---|
| Mistral Medium 3.1 | 0.777 | LLMs |
| Mistral Large 3 | 0.76 | LLMs |
| Anthropic Claude Fable 5 | 0.615 | LLMs |
| Anthropic Claude Opus 4.8 | 0.614 | LLMs |
| Anthropic Claude Sonnet 5 | 0.611 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.58 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.576 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.459 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.458 | LLMs |
| MoonshotAI Kimi K2.5 | 0.444 | LLMs |
| AI21 Jamba Large 1.7 | 0.427 | LLMs |
| z.ai GLM-5 | 0.383 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.372 | Open-weight LLMs |
| DeepSeek V3.2 | 0.368 | LLMs |
| Cohere Command A Translate | 0.358 | LLMs |
| Google Gemma 4 | 0.344 | Open-weight LLMs |
| Google Gemini 3.5 Flash | 0.343 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.34 | LLMs |
| OpenAI GPT-4.1 | 0.325 | LLMs |
| OpenAI GPT-5.6 Sol | 0.323 | LLMs |
| OpenAI GPT-5.5 | 0.322 | LLMs |
| OpenAI GPT-5.6 Luna | 0.322 | LLMs |
| OpenAI GPT-5.6 Terra | 0.321 | LLMs |
| Cohere Tiny Aya | 0.308 | LLMs |
| MiniMax M2.5 | 0.262 | LLMs |
The tokenizers produce a median 0.37 tokens per character of Arabic (MSA), 1.6 times their count for English. For the two solutions that meet every Standard requirement, the tokenizer of Anthropic Claude Fable 5 produces 0.615 tokens per character and that of OpenAI GPT-5.5 0.322, and the solutions cost $32.52 and $10.27 per million characters. The cost does not follow the tokens alone: at 0.611 tokens per character, the one-step solution on Anthropic Claude Sonnet 5 costs $6.52 per million characters.
| 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 | 58 |
| Standard | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | 58 |
| Advanced | No solution | — |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.2 | MT engines |
| Google Cloud Advanced Translation API | 0.3 | MT engines |
| Pangeanic PangeaMT | 0.6 | MT engines |
| Google Translation LLM | 0.7 | LLM-based MT engines |
| DeepL Next Gen | 0.9 | LLM-based MT engines |
| ModernMT Enterprise Edition | 1 | MT engines |
| Tarjama NMT | 1 | MT engines |
| Translated Lara 2 | 1.2 | LLM-based MT engines |
| Amazon Translate | 1.3 | MT engines |
| OpenAI gpt-oss-120b | 2 | Open-weight LLMs |
| SYSTRAN PNMT API | 2.2 | MT engines |
| Yandex Cloud Translate v2 | 2.6 | MT engines |
| Oracle Cloud Language Service | 4.8 | MT engines |
| Cohere Tiny Aya | 12 | LLMs |
| OpenAI GPT-5.6 Luna | 13 | LLMs |
| OpenAI GPT-4.1 | 18 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 18 | LLMs |
| Alibaba Machine Translation API | 22 | MT engines |
| OpenAI GPT-5.6 Terra | 30 | LLMs |
| Cohere Command A Translate | 32 | LLMs |
| Mistral Medium 3.1 | 34 | LLMs |
| ALL-Lab | 35 | MT engines |
| Swiss AI Apertus-8B-2509 | 37 | Open-weight LLMs |
| Anthropic Claude Haiku 4.5 | 40 | LLMs |
| Meta Llama 4 Maverick | 40 | Open-weight LLMs |
| MiniMax M2.5 | 43 | LLMs |
| iFlyTek MT | 47 | MT engines |
| xAI Grok 4.20 (reasoning) | 50 | LLMs |
| Anthropic Claude Sonnet 5 | 51 | LLMs |
| Mistral Large 3 | 52 | LLMs |
| AI21 Jamba Large 1.7 | 54 | LLMs |
| DeepSeek V3.2 | 55 | LLMs |
| OpenAI GPT-5.6 Sol | 56 | LLMs |
| Anthropic Claude Fable 5 | 58 | LLMs |
| Google Gemini 3.5 Flash | 62 | LLMs |
| Anthropic Claude Opus 4.8 | 62 | LLMs |
| OpenAI GPT-5.5 | 67 | LLMs |
| Google Gemini 3.1 Pro Preview | 72 | LLMs |
| AppTek NMT | 75 | MT engines |
| Anthropic Claude Sonnet 4.6 | 87 | LLMs |
| Alibaba Qwen3-235B-A22B | 93 | Open-weight LLMs |
| z.ai GLM-5 | 160 | LLMs |
| Google Gemma 4 | 197 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 297 | 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 Arabic (MSA).
The real-time models in Arabic (MSA) are seven MT engines, three LLM-based MT engines and the open-weight OpenAI gpt-oss-120b. None of the three real-time solutions on the shortlist meet general translation quality or tag handling. The two solutions that meet every Standard requirement take 58 and 67 seconds per 1,000 words, and their three-step versions with the gender-neutrality agent 239 and 285 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 | three-step, with the gender-neutrality agent | 99.1% | ✗ | $1,005 | 285 |
| Anthropic Claude Fable 5 | three-step, with the gender-neutrality agent | 98.2% | ✗ | $1,047 | 239 |
For Arabic (MSA) we check the formal register and gender neutrality through neutral rephrasing, with no default gender. Masculine default and double (slash) forms are the other options in the language. Translations from eight of the ten shortlisted LLMs have 100% of the sentences in the formal register, and those from Google Gemini 3.1 Pro Preview and Cohere Command A Translate 97.9%. Translations from the MT engine ALL-Lab and the LLM-based MT engines DeepL Next Gen and Google Translation LLM have 95.7% to 97.9%.
Leave feedback