The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Marathi 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 Marathi. |
| Tone of voice | The formal register (आपण) 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 Marathi, 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.
Two three-step solutions (translation + LQA + repair) meet every Standard requirement in Marathi, and neither leads on errors, cost and time at once. The one on OpenAI GPT-5.6 Sol has 6.2 total errors per 1,000 words against 11.3 for the one on Google Gemini 3.1 Pro Preview, and costs $547 against $553 per million characters. The one on Google Gemini 3.1 Pro Preview takes 274 seconds per 1,000 words against 361. Optimize for selects OpenAI GPT-5.6 Sol for quality and cost, and Google Gemini 3.1 Pro Preview for latency.
Under the Standard requirements, a cost cap below the $553 per million characters of the one on Google Gemini 3.1 Pro Preview leaves at most the one on OpenAI GPT-5.6 Sol, and a time limit below the 361 seconds per 1,000 words of the one on OpenAI GPT-5.6 Sol at most the one on Google Gemini 3.1 Pro Preview. With a limit of one step, only the one on OpenAI GPT-5.6 Sol meets even the Basic requirements, at $8.20 per million characters and 175 seconds per 1,000 words. None of the real-time (<2 sec) solutions on the shortlist meet them.
For the Advanced requirements, the gender-neutrality agent works with both three-step solutions (translation + LQA + repair). On OpenAI GPT-5.6 Sol it raises the cost from $547 to $902 per million characters, and total errors fall from 6.2 to 5.9 per 1,000 words. The time goes from 361 to 362 seconds per 1,000 words. On Google Gemini 3.1 Pro Preview the cost goes from $553 to $946 per million characters, and total errors rise from 11.3 to 11.5 per 1,000 words. Its time goes from 274 to 277 seconds per 1,000 words.
Google Gemini 3.1 Pro Preview
three-step
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
three-step
OpenAI GPT-5.6 Sol
three-step
Google Gemini 3.1 Pro Preview
three-step
OpenAI GPT-5.6 Sol
three-step with the gender-neutrality agent
OpenAI GPT-5.6 Sol
three-step with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
three-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 | OpenAI GPT-5.6 Sol | 17.6 | $8.20 | 175 |
| Standard | Tier 2: two or more solutions, none simple | ☆☆ | three-step | OpenAI GPT-5.6 Sol | 6.2 | $547 | 361 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | with the gender-neutrality agent | OpenAI GPT-5.6 Sol | 5.9 | $902 | 362 |
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 OpenAI GPT-5.6 Sol for cost and latency, and the three-step solution (translation + LQA + repair) on Google Gemini 3.1 Pro Preview for quality, with 11.3 total errors per 1,000 words. The Standard requirements remove the one-step solution, which misses tag handling, so both solutions that meet them take three steps. The Advanced requirements add the gender-neutrality agent to both and keep the selections of the Standard ones.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.13 | 54.4 | Misses a Standard requirement |
| Mistral Large 3 | 1.47 | 78.4 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.2 | 44.5 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 3.21 | 72.2 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.2 | 30.3 | Misses a Standard requirement |
| Anthropic Claude Sonnet 5 | 7.86 | 37.9 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 8.2 | 16.2 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 11.64 | 26.4 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 53.7 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 75.8 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 39.74 | 23.1 | Misses a Standard requirement |
| Google Translation LLM | 42.87 | 112.4 | Misses a Standard requirement |
Among the shortlisted one-step solutions in Marathi, paying more than $8.20 per million characters does not buy fewer weighted errors. The one on OpenAI GPT-5.6 Sol has the fewest, 16.2 weighted errors per 1,000 words, at $8.20 per million characters. The five dearer ones cost $11.64 to $42.87 per million characters and have 23.1 to 112.4 weighted errors per 1,000 words. The cheapest, on the open-weight Google Gemma 4, costs $0.13 per million characters and has 54.4 weighted errors per 1,000 words, fewer than four of the others.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 49.1 |
| Tag handling | 2 |
| Gender neutrality | 1.6 |
| Tone of voice | 0.1 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
General translation quality and tag handling split the one-step solutions. Only the one on OpenAI GPT-5.6 Sol meets the first. Only those on Google Gemini 3.1 Pro Preview and Translated Lara 2 get every tag right, with 30.3 and 53.7 weighted errors per 1,000 words. The two-step solution (translation + post-edit) on OpenAI GPT-5.6 Sol gets every tag right but misses formatting, and the one on Google Gemini 3.1 Pro Preview has 24.4 weighted errors per 1,000 words. Three-step solutions (translation + LQA + repair) meet both.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Marathi | Class |
|---|---|---|
| Google Translation LLM | 112.4 | LLM-based MT engines |
| Mistral Large 3 | 78.4 | LLMs |
| DeepL Next Gen | 75.8 | LLM-based MT engines |
| OpenAI GPT-4.1 | 72.2 | LLMs |
| Google Gemma 4 | 54.4 | Open-weight LLMs |
| Translated Lara 2 | 53.7 | LLM-based MT engines |
| Google Gemini 3.5 Flash | 44.5 | LLMs |
| Anthropic Claude Sonnet 5 | 37.9 | LLMs |
| Google Gemini 3.1 Pro Preview | 30.3 | LLMs |
| OpenAI GPT-5.5 | 26.4 | LLMs |
| Anthropic Claude Fable 5 | 23.1 | LLMs |
| OpenAI GPT-5.6 Sol | 16.2 | LLMs |
Translations into Marathi from the three LLM-based MT engines on the shortlist, Translated Lara 2, DeepL Next Gen and Google Translation LLM, have 53.7 to 112.4 weighted errors per 1,000 words. Only those from Google Translation LLM have critical errors, 10 of them. Translations from OpenAI GPT-4.1 and Mistral Large 3, the two shortlisted LLMs below our language-support criteria, have 72.2 and 78.4 weighted errors per 1,000 words. The open-weight Google Gemma 4 meets the criteria, with 54.4 weighted errors per 1,000 words.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Anthropic Claude Opus 4.8 | 0.696 | 9 | 1.89 | 14 | No | Not shortlisted |
| DeepSeek V3.2 | 0.695 | 11 | 1.84 | 10 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.694 | 13 | 1.94 | 15 | No | Not shortlisted |
| z.ai GLM-5 | 0.693 | 14 | 1.94 | 16 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.687 | 17 | 2.06 | 18 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.681 | 18 | 1.94 | 17 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.679 | 19 | 2.61 | 27 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.677 | 20 | 2.29 | 25 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.677 | 21 | 2.12 | 19 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.675 | 22 | 1.89 | 13 | No | Not shortlisted |
| MiniMax M2.5 | 0.675 | 23 | 2.18 | 20 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.673 | 24 | 2.28 | 24 | No | Not shortlisted |
| Microsoft Translator | 0.668 | 25 | 2.24 | 22 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.666 | 26 | 2.36 | 26 | No | Not shortlisted |
| Cohere Command A Translate | 0.663 | 27 | 2.28 | 23 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.639 | 28 | 3.26 | 29 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.629 | 29 | 2.88 | 28 | No | Not shortlisted |
| Cohere Tiny Aya | 0.602 | 30 | 4.08 | 30 | No | Not shortlisted |
| AppTek NMT | 0.529 | 31 | 5.84 | 31 | No | Not shortlisted |
| Amazon Translate | 0.467 | 32 | 6.80 | 32 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.416 | 33 | 8.88 | 33 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.399 | 34 | 10.87 | 34 | No | Not shortlisted |
| OpenAI GPT-5.6 Sol | 0.707 | 1 | 1.64 | 2 | Yes | Top 8 by both metrics |
| Google Gemini 3.5 Flash | 0.706 | 2 | 1.85 | 11 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | 0.705 | 3 | 1.72 | 5 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.705 | 4 | 1.85 | 12 | Yes | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | 0.702 | 5 | 1.67 | 3 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.702 | 6 | 1.62 | 1 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.5 | 0.698 | 7 | 1.80 | 7 | Yes | Top 8 by both metrics |
| DeepL Next Gen | 0.696 | 8 | 1.82 | 8 | Yes | Top 8 by both metrics |
| Translated Lara 2 | 0.696 | 10 | 1.82 | 9 | Yes | Real-time place |
| Mistral Large 3 | 0.694 | 12 | 1.72 | 4 | Yes | Top 8 by MetricX |
| OpenAI GPT-4.1 | 0.692 | 15 | 1.77 | 6 | Yes | Top 8 by MetricX |
| Google Translation LLM | 0.691 | 16 | 2.20 | 21 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| OpenAI GPT-5.6 Sol | Adjusted LLM | 1 | 0.707 | 2 | 1.64 | Top 8 by both metrics |
| Google Gemini 3.5 Flash | Adjusted LLM | 2 | 0.706 | 11 | 1.85 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 3 | 0.705 | 5 | 1.72 | Top 8 by both metrics |
| Google Gemma 4 | Adjusted LLM | 4 | 0.705 | 12 | 1.85 | Top 8 by COMET-Kiwi |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 5 | 0.702 | 3 | 1.67 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 6 | 0.702 | 1 | 1.62 | Top 8 by both metrics |
| OpenAI GPT-5.5 | Adjusted LLM | 7 | 0.698 | 7 | 1.80 | Top 8 by both metrics |
| DeepL Next Gen | Off-the-shelf NMT | 8 | 0.696 | 8 | 1.82 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 9 | 0.696 | 14 | 1.89 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 10 | 0.696 | 9 | 1.82 | Real-time place |
| DeepSeek V3.2 | Adjusted LLM | 11 | 0.695 | 10 | 1.84 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 12 | 0.694 | 4 | 1.72 | Top 8 by MetricX |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 13 | 0.694 | 15 | 1.94 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 14 | 0.693 | 16 | 1.94 | Not shortlisted |
| OpenAI GPT-4.1 | Adjusted LLM | 15 | 0.692 | 6 | 1.77 | Top 8 by MetricX |
| Google Translation LLM | Off-the-shelf NMT | 16 | 0.691 | 21 | 2.20 | Real-time place |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 17 | 0.687 | 18 | 2.06 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 18 | 0.681 | 17 | 1.94 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 19 | 0.679 | 27 | 2.61 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 20 | 0.677 | 25 | 2.29 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Off-the-shelf LLM | 21 | 0.677 | 19 | 2.12 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 22 | 0.675 | 13 | 1.89 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 23 | 0.675 | 20 | 2.18 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 24 | 0.673 | 24 | 2.28 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 25 | 0.668 | 22 | 2.24 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 26 | 0.666 | 26 | 2.36 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 27 | 0.663 | 23 | 2.28 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 28 | 0.639 | 29 | 3.26 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 29 | 0.629 | 28 | 2.88 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 30 | 0.602 | 30 | 4.08 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 31 | 0.529 | 31 | 5.84 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 32 | 0.467 | 32 | 6.80 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 33 | 0.416 | 33 | 8.88 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 34 | 0.399 | 34 | 10.87 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.699 | — | 1.60 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.700 | — | 1.77 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.676 | — | 1.95 | 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 34 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.
COMET-Kiwi ranks Google Gemini 3.5 Flash second and Google Gemma 4 fourth in Marathi, and MetricX ranks them 11th and 12th, so each is on the shortlist through COMET-Kiwi alone. By Intento LQA, the one-step solutions on them have 44.5 and 54.4 weighted errors per 1,000 words. The one on Anthropic Claude Fable 5, first on MetricX, has 23.1 weighted errors per 1,000 words, against 16.2 for the one on OpenAI GPT-5.6 Sol. In the tail, DeepSeek V3.2 sits nearest the cut on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.6 Sol, one-step LLM adjusted for Standard requirements | $8.20 |
| Standard | OpenAI GPT-5.6 Sol, three-step | $547 |
| Advanced | OpenAI GPT-5.6 Sol, three-step with the gender-neutrality agent | $902 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.13 | Open-weight LLMs |
| Mistral Large 3 | $1.47 | LLMs |
| Google Gemini 3.5 Flash | $3.2 | LLMs |
| OpenAI GPT-4.1 | $3.21 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.2 | LLMs |
| Anthropic Claude Sonnet 5 | $7.86 | LLMs |
| OpenAI GPT-5.6 Sol | $8.2 | LLMs |
| OpenAI GPT-5.5 | $11.64 | LLMs |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $39.74 | LLMs |
| Google Translation LLM | $42.87 | LLM-based MT engines |
| Model | Tokens per character | Class |
|---|---|---|
| Mistral Large 3 | 0.888 | LLMs |
| Mistral Medium 3.1 | 0.884 | LLMs |
| AI21 Jamba Large 1.7 | 0.849 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.808 | Open-weight LLMs |
| z.ai GLM-5 | 0.792 | LLMs |
| Anthropic Claude Sonnet 5 | 0.691 | LLMs |
| Anthropic Claude Fable 5 | 0.69 | LLMs |
| Cohere Command A Translate | 0.69 | LLMs |
| Anthropic Claude Opus 4.8 | 0.688 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.659 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.653 | LLMs |
| DeepSeek V3.2 | 0.551 | LLMs |
| MoonshotAI Kimi K2.5 | 0.529 | LLMs |
| MiniMax M2.5 | 0.447 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.41 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.402 | LLMs |
| OpenAI GPT-4.1 | 0.353 | LLMs |
| OpenAI GPT-5.5 | 0.351 | LLMs |
| OpenAI GPT-5.6 Sol | 0.349 | LLMs |
| OpenAI GPT-5.6 Luna | 0.348 | LLMs |
| OpenAI GPT-5.6 Terra | 0.348 | LLMs |
| Cohere Tiny Aya | 0.298 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.294 | LLMs |
| Google Gemma 4 | 0.291 | Open-weight LLMs |
| Google Gemini 3.5 Flash | 0.29 | LLMs |
The tokenizers produce a median 0.53 tokens per character of the Marathi translation, and a median 2.0 times their count for English. Tokens alone do not set the cost. The tokenizer of Anthropic Claude Sonnet 5 produces 0.691 tokens per character and that of OpenAI GPT-5.6 Sol 0.349, yet the one-step solutions on them cost $7.86 and $8.20 per million characters. The steps weigh more. On OpenAI GPT-5.6 Sol, meeting the Basic requirements costs $8.20 per million characters, the Standard ones $547 and the Advanced ones $902.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | OpenAI GPT-5.6 Sol, one-step LLM adjusted for Standard requirements | 175 |
| Standard | Google Gemini 3.1 Pro Preview, three-step | 274 |
| Advanced | Google Gemini 3.1 Pro Preview, three-step with the gender-neutrality agent | 277 |
| 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 | 0.4 | MT engines |
| DeepL Next Gen | 0.9 | LLM-based MT engines |
| Amazon Translate | 1.3 | MT engines |
| Translated Lara 2 | 2.1 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 3 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 4.1 | MT engines |
| Google Translation LLM | 12 | LLM-based MT engines |
| Google Gemini 3.5 Flash | 17 | LLMs |
| OpenAI GPT-5.6 Luna | 17 | LLMs |
| Cohere Tiny Aya | 18 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 18 | LLMs |
| OpenAI GPT-4.1 | 25 | LLMs |
| Alibaba Machine Translation API | 28 | MT engines |
| Swiss AI Apertus-8B-2509 | 49 | Open-weight LLMs |
| OpenAI GPT-5.6 Terra | 55 | LLMs |
| xAI Grok 4.20 (reasoning) | 56 | LLMs |
| MiniMax M2.5 | 56 | LLMs |
| Anthropic Claude Haiku 4.5 | 59 | LLMs |
| Mistral Medium 3.1 | 63 | LLMs |
| AI21 Jamba Large 1.7 | 71 | LLMs |
| DeepSeek V3.2 | 72 | LLMs |
| Anthropic Claude Sonnet 5 | 78 | LLMs |
| Cohere Command A Translate | 79 | LLMs |
| Mistral Large 3 | 82 | LLMs |
| Anthropic Claude Fable 5 | 93 | LLMs |
| Google Gemini 3.1 Pro Preview | 97 | LLMs |
| Anthropic Claude Opus 4.8 | 108 | LLMs |
| AppTek NMT | 114 | MT engines |
| Anthropic Claude Sonnet 4.6 | 120 | LLMs |
| OpenAI GPT-5.5 | 121 | LLMs |
| Google Gemma 4 | 148 | Open-weight LLMs |
| OpenAI GPT-5.6 Sol | 175 | LLMs |
| Alibaba Qwen3-235B-A22B | 283 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 325 | LLMs |
| z.ai GLM-5 | 366 | 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 Marathi.
Of the eight real-time models in Marathi, four are MT engines, three are LLM-based MT engines (DeepL Next Gen, Google Translation LLM and Translated Lara 2) and one is the open-weight LLM OpenAI gpt-oss-120b. Translations from the real-time solutions on the shortlist have 53.7 or more weighted errors per 1,000 words. The fastest solutions that meet the Basic, Standard and Advanced requirements take 175, 274 and 277 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, with the gender-neutrality agent | 100% | ✓ | $902 | 362 |
| Google Gemini 3.1 Pro Preview | three-step, with the gender-neutrality agent | 100% | ✓ | $946 | 277 |
For Marathi we check the formal register, with आपण, and gender-neutral wording through neutral nouns and neutral rephrasing, never a default gender. The other options are a masculine default and explicit double forms. Translations from the one-step solutions on DeepL Next Gen and Google Translation LLM have 97.87% and 93.6% of the sentences in the formal register, and those from the other ten have 100%. With the gender-neutrality agent, 100% of the sentences about people in translations from the two best solutions are gender-neutral.
For the 12 shortlisted solutions with an Intento LQA result in Marathi, the median share of sentences on which the three LQA judges agree is 93.8%. With agreement that high, the Intento LQA error counts on this page are reliable. Both pairs of automatic metrics agree above their thresholds, the similarity pair by a wide margin and the quality-estimation pair by a narrow one, and the result is not borderline.
Leave feedback