The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Korean 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 (하십시오체, 합니다체, with honorifics) 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 Korean, 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.
Three solutions meet every Standard requirement in Korean, all three-step (translation + LQA + repair). The one on Google Gemini 3.1 Pro Preview has the fewest total errors, 13.8 per 1,000 words, and the lowest cost. The one on Google Gemini 3.5 Flash takes the shortest time, with 19.1 total errors per 1,000 words. The one on Google Gemma 4 has 14.9 total errors per 1,000 words and is the dearest and slowest of the three. Optimizing for quality or cost selects Google Gemini 3.1 Pro Preview, and for latency Google Gemini 3.5 Flash.
None of the one-step or two-step solutions (translation + post-edit) in Korean meet even the Basic requirements, the real-time ones included. So every solution that meets them takes three steps (translation + LQA + repair), at $577 per million characters or more and 118 seconds per 1,000 words or more. Of the solutions that meet every Standard requirement, only the one on Google Gemini 3.5 Flash takes less than 293 seconds per 1,000 words.
Two three-step solutions (translation + LQA + repair) meet every Advanced requirement in Korean, and only the one on Google Gemma 4 does so with the gender-neutrality agent. The agent raises its cost from $594 to $1,000 per million characters and its time from 505 to 596 seconds per 1,000 words. Its total errors fall from 14.9 to 11.9 per 1,000 words. The one on Google Gemini 3.1 Pro Preview meets them at $577 per million characters and 293 seconds per 1,000 words.
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.5 Flash
three-step
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.5 Flash
three-step
Google Gemma 4
three-step with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.1 Pro Preview
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 | ☆×4 | three-step | Google Gemini 3.1 Pro Preview | 13.8 | $577 | 293 |
| Standard | Tier 2: two or more solutions, none simple | ☆☆☆ | three-step | Google Gemini 3.1 Pro Preview | 13.8 | $577 | 293 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | three-step | Google Gemini 3.1 Pro Preview | 13.8 | $577 | 293 |
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.
Four three-step solutions (translation + LQA + repair) meet every Basic requirement in Korean. The Standard requirements remove the one on Anthropic Claude Sonnet 4.6, with 99.4% of the tags right. The Advanced requirements keep two, on Google Gemini 3.1 Pro Preview and on Google Gemma 4 with the gender-neutrality agent. Optimizing for quality or cost selects Google Gemini 3.1 Pro Preview and for latency Google Gemini 3.5 Flash under the Basic and Standard requirements, and for quality Google Gemma 4 under the Advanced ones.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.13 | 40.5 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.08 | 36.6 | Misses a Standard requirement |
| OpenAI GPT-5.6 Terra | 4.03 | 21.4 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.12 | 27 | Misses a Standard requirement |
| Anthropic Claude Sonnet 5 | 6.11 | 35.6 | Misses a Standard requirement |
| Anthropic Claude Sonnet 4.6 | 8.34 | 32 | Misses a Standard requirement |
| Amazon Translate | 15 | 100.2 | Misses a Standard requirement |
| Anthropic Claude Opus 4.8 | 15.42 | 23.7 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 41.2 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 32.6 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 30.3 | 24.7 | Misses a Standard requirement |
In one step in Korean, no solution that costs more than $4.03 per million characters has fewer weighted errors than the one at that price. That one, on OpenAI GPT-5.6 Terra, has 21.4 weighted errors per 1,000 words, the fewest in one step. The two cheaper ones, on Google Gemma 4 and Google Gemini 3.5 Flash, have 40.5 and 36.6 weighted errors per 1,000 words. The three-step solution (translation + LQA + repair) on Google Gemini 3.1 Pro Preview has 13.8 weighted errors per 1,000 words at $577 per million characters.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 34.1 |
| Tag handling | 1.7 |
| Terminology | 0 |
| Tone of voice | 0 |
| Gender neutrality | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
Without the gender-neutrality agent, the six three-step solutions (translation + LQA + repair) in Korean have 13.8 to 19.1 weighted errors per 1,000 words and no critical error. Four of them meet general translation quality, and those on Anthropic Claude Fable 5 and Anthropic Claude Sonnet 5 miss it on major errors alone, with 4 and 5 major errors.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Korean | Class |
|---|---|---|
| Amazon Translate | 100.2 | MT engines |
| Cohere Command A Translate | 48.8 | LLMs |
| Translated Lara 2 | 41.2 | LLM-based MT engines |
| Google Gemma 4 | 40.5 | Open-weight LLMs |
| Google Gemini 3.5 Flash | 36.6 | LLMs |
| Anthropic Claude Sonnet 5 | 35.6 | LLMs |
| DeepL Next Gen | 32.6 | LLM-based MT engines |
| Anthropic Claude Sonnet 4.6 | 32 | LLMs |
| Google Gemini 3.1 Pro Preview | 27 | LLMs |
| Anthropic Claude Fable 5 | 24.7 | LLMs |
| Anthropic Claude Opus 4.8 | 23.7 | LLMs |
| OpenAI GPT-5.6 Terra | 21.4 | LLMs |
No translation into Korean: Widn AI Tower Anthill (garbage or near-empty, cut off part-way, wrong language, wrong script), Widn AI Tower Vesuvius (cut off part-way).
In one step, translations from the nine shortlisted LLMs have 21.4 to 48.8 weighted errors per 1,000 words in Korean. Translations from the LLM-based MT engines DeepL Next Gen and Translated Lara 2 have 32.6 and 41.2 weighted errors per 1,000 words, and those from Amazon Translate 100.2, with 3 critical errors. Translations from all nine LLMs meet formatting. None of those from the two LLM-based MT engines and Amazon Translate meet it. All nine LLMs, the open-weight Google Gemma 4 included, support Korean by our measured criteria.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| MoonshotAI Kimi K2.5 | 0.806 | 10 | 1.16 | 11 | No | Not shortlisted |
| OpenAI GPT-5.5 | 0.805 | 11 | 1.21 | 16 | No | Not shortlisted |
| z.ai GLM-5 | 0.803 | 13 | 1.15 | 9 | No | Not shortlisted |
| Mistral Large 3 | 0.803 | 14 | 1.15 | 10 | No | Not shortlisted |
| DeepSeek V3.2 | 0.802 | 15 | 1.23 | 20 | No | Not shortlisted |
| Widn AI Tower Sugarloaf | 0.802 | 16 | 1.21 | 14 | No | Not shortlisted |
| OpenAI GPT-4.1 | 0.799 | 17 | 1.18 | 13 | No | Not shortlisted |
| Microsoft Translator | 0.797 | 20 | 1.17 | 12 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.797 | 21 | 1.35 | 22 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.796 | 22 | 1.26 | 21 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.794 | 23 | 1.23 | 18 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.791 | 24 | 1.37 | 23 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.790 | 25 | 1.37 | 24 | No | Not shortlisted |
| Cohere Tiny Aya | 0.789 | 26 | 1.48 | 26 | No | Not shortlisted |
| MiniMax M2.5 | 0.783 | 27 | 1.64 | 30 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.779 | 28 | 1.55 | 28 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.778 | 29 | 1.49 | 27 | No | Not shortlisted |
| Google Translation LLM | 0.777 | 30 | 1.87 | 33 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.773 | 31 | 1.61 | 29 | No | Not shortlisted |
| Naver Papago NMT Commercial | 0.769 | 32 | 1.78 | 31 | No | Not shortlisted |
| Supertext LLM | 0.760 | 33 | 1.78 | 32 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.759 | 34 | 1.88 | 34 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.759 | 35 | 2.04 | 36 | No | Not shortlisted |
| Youdao Translation Service API | 0.759 | 36 | 1.92 | 35 | No | Not shortlisted |
| iFlyTek MT | 0.737 | 37 | 2.19 | 37 | No | Not shortlisted |
| Oracle Cloud Language Service | 0.725 | 38 | 2.42 | 38 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.721 | 39 | 2.43 | 39 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.698 | 40 | 3.11 | 40 | No | Not shortlisted |
| AppTek NMT | 0.577 | 41 | 6.09 | 41 | No | Not shortlisted |
| DeepL Next Gen | 0.816 | 1 | 0.99 | 1 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | 0.813 | 2 | 1.10 | 7 | Yes | Top 8 by both metrics |
| Cohere Command A Translate | 0.811 | 3 | 1.07 | 5 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.811 | 4 | 1.03 | 3 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 4.6 | 0.810 | 5 | 1.03 | 2 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.810 | 6 | 1.06 | 4 | Yes | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | 0.809 | 7 | 1.22 | 17 | Yes | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.6 Terra | 0.808 | 8 | 1.21 | 15 | Yes | Top 8 by COMET-Kiwi |
| Google Gemma 4 | 0.807 | 9 | 1.08 | 6 | Yes | Top 8 by MetricX |
| Google Gemini 3.5 Flash | 0.804 | 12 | 1.11 | 8 | Yes | Top 8 by MetricX |
| Translated Lara 2 | 0.798 | 18 | 1.38 | 25 | Yes | Real-time place |
| Amazon Translate | 0.798 | 19 | 1.23 | 19 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| DeepL Next Gen | Adjusted NMT | 1 | 0.816 | 1 | 0.99 | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 2 | 0.813 | 7 | 1.10 | Top 8 by both metrics |
| Cohere Command A Translate | Adjusted LLM | 3 | 0.811 | 5 | 1.07 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 4 | 0.811 | 3 | 1.03 | Top 8 by both metrics |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 5 | 0.810 | 2 | 1.03 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 6 | 0.810 | 4 | 1.06 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 7 | 0.809 | 17 | 1.22 | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.6 Terra | Adjusted LLM | 8 | 0.808 | 15 | 1.21 | Top 8 by COMET-Kiwi |
| Google Gemma 4 | Adjusted LLM | 9 | 0.807 | 6 | 1.08 | Top 8 by MetricX |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 10 | 0.806 | 11 | 1.16 | Not shortlisted |
| OpenAI GPT-5.5 | Adjusted LLM | 11 | 0.805 | 16 | 1.21 | Not shortlisted |
| Google Gemini 3.5 Flash | Adjusted LLM | 12 | 0.804 | 8 | 1.11 | Top 8 by MetricX |
| z.ai GLM-5 | Adjusted LLM | 13 | 0.803 | 9 | 1.15 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 14 | 0.803 | 10 | 1.15 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 15 | 0.802 | 20 | 1.23 | Not shortlisted |
| Widn AI Tower Sugarloaf | Adjusted NMT | 16 | 0.802 | 14 | 1.21 | Not shortlisted |
| OpenAI GPT-4.1 | Adjusted LLM | 17 | 0.799 | 13 | 1.18 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 18 | 0.798 | 25 | 1.38 | Real-time place |
| Amazon Translate | Off-the-shelf NMT | 19 | 0.798 | 19 | 1.23 | Real-time place |
| Microsoft Translator | Off-the-shelf NMT | 20 | 0.797 | 12 | 1.17 | Not shortlisted |
| Mistral Medium 3.1 | Off-the-shelf LLM | 21 | 0.797 | 22 | 1.35 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 22 | 0.796 | 21 | 1.26 | Not shortlisted |
| OpenAI gpt-oss-120b | Off-the-shelf LLM | 23 | 0.794 | 18 | 1.23 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 24 | 0.791 | 23 | 1.37 | Not shortlisted |
| ModernMT Enterprise Edition | Adjusted NMT | 25 | 0.790 | 24 | 1.37 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 26 | 0.789 | 26 | 1.48 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 27 | 0.783 | 30 | 1.64 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 28 | 0.779 | 28 | 1.55 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 29 | 0.778 | 27 | 1.49 | Not shortlisted |
| Google Translation LLM | Off-the-shelf NMT | 30 | 0.777 | 33 | 1.87 | Not shortlisted |
| SYSTRAN PNMT API | Adjusted NMT | 31 | 0.773 | 29 | 1.61 | Not shortlisted |
| Naver Papago NMT Commercial | Adjusted NMT | 32 | 0.769 | 31 | 1.78 | Not shortlisted |
| Supertext LLM | Off-the-shelf NMT | 33 | 0.760 | 32 | 1.78 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 34 | 0.759 | 34 | 1.88 | Not shortlisted |
| Google Cloud Advanced Translation API | Adjusted NMT | 35 | 0.759 | 36 | 2.04 | Not shortlisted |
| Youdao Translation Service API | Off-the-shelf NMT | 36 | 0.759 | 35 | 1.92 | Not shortlisted |
| iFlyTek MT | Off-the-shelf NMT | 37 | 0.737 | 37 | 2.19 | Not shortlisted |
| Oracle Cloud Language Service | Off-the-shelf NMT | 38 | 0.725 | 38 | 2.42 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 39 | 0.721 | 39 | 2.43 | Not shortlisted |
| Yandex Cloud Translate v2 | Adjusted NMT | 40 | 0.698 | 40 | 3.11 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 41 | 0.577 | 41 | 6.09 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.807 | — | 1.25 | Same model family as OpenAI GPT-5.6 Terra |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.807 | — | 1.13 | Same model family as OpenAI GPT-5.6 Terra |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.779 | — | 1.75 | 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.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | Google Gemini 3.1 Pro Preview, three-step | $577 |
| Standard | Google Gemini 3.1 Pro Preview, three-step | $577 |
| Advanced | Google Gemini 3.1 Pro Preview, three-step | $577 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.13 | Open-weight LLMs |
| Google Gemini 3.5 Flash | $3.08 | LLMs |
| OpenAI GPT-5.6 Terra | $4.03 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.12 | LLMs |
| Anthropic Claude Sonnet 5 | $6.11 | LLMs |
| Anthropic Claude Sonnet 4.6 | $8.34 | LLMs |
| Amazon Translate | $15 | MT engines |
| Anthropic Claude Opus 4.8 | $15.42 | LLMs |
| 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 |
|---|---|---|
| Mistral Medium 3.1 | 0.83 | LLMs |
| Mistral Large 3 | 0.825 | LLMs |
| Anthropic Claude Opus 4.8 | 0.807 | LLMs |
| Anthropic Claude Fable 5 | 0.803 | LLMs |
| Anthropic Claude Sonnet 5 | 0.8 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.759 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.746 | LLMs |
| AI21 Jamba Large 1.7 | 0.745 | LLMs |
| MoonshotAI Kimi K2.5 | 0.596 | LLMs |
| z.ai GLM-5 | 0.583 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.53 | Open-weight LLMs |
| Widn AI Tower Sugarloaf | 0.516 | LLM-based MT engines |
| DeepSeek V3.2 | 0.515 | LLMs |
| Cohere Command A Translate | 0.464 | LLMs |
| OpenAI GPT-4.1 | 0.464 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.457 | LLMs |
| Google Gemini 3.5 Flash | 0.457 | LLMs |
| OpenAI GPT-5.6 Sol | 0.456 | LLMs |
| OpenAI GPT-5.6 Terra | 0.454 | LLMs |
| OpenAI GPT-5.5 | 0.453 | LLMs |
| Google Gemma 4 | 0.452 | Open-weight LLMs |
| OpenAI GPT-5.6 Luna | 0.452 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.448 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.446 | LLMs |
| Cohere Tiny Aya | 0.388 | LLMs |
| MiniMax M2.5 | 0.321 | LLMs |
The tokenizers of the four Anthropic Claude models on the shortlist produce 0.759 to 0.807 tokens per character of the Korean translation, and those of the four Google and OpenAI models 0.452 to 0.457. On one model, the cost follows the steps. The one-step solution on Google Gemini 3.1 Pro Preview costs $4.12 per million characters, and the three-step one (translation + LQA + repair) $577 per million characters. It is the cheapest solution that meets the Basic, the Standard and the Advanced requirements.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | Google Gemini 3.5 Flash, three-step | 118 |
| Standard | Google Gemini 3.5 Flash, three-step | 118 |
| Advanced | Google Gemini 3.1 Pro Preview, three-step | 293 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.2 | MT engines |
| Google Cloud Advanced Translation API | 0.2 | MT engines |
| DeepL Next Gen | 0.6 | LLM-based MT engines |
| Google Translation LLM | 0.9 | LLM-based MT engines |
| ModernMT Enterprise Edition | 1 | MT engines |
| Translated Lara 2 | 1.2 | LLM-based MT engines |
| Amazon Translate | 1.5 | MT engines |
| OpenAI gpt-oss-120b | 1.9 | Open-weight LLMs |
| SYSTRAN PNMT API | 2 | MT engines |
| Yandex Cloud Translate v2 | 2.6 | MT engines |
| Oracle Cloud Language Service | 5 | MT engines |
| xAI Grok 4.20 (non-reasoning) | 12 | LLMs |
| OpenAI GPT-5.6 Luna | 13 | LLMs |
| Google Gemini 3.5 Flash | 14 | LLMs |
| Supertext LLM | 15 | LLM-based MT engines |
| Cohere Tiny Aya | 16 | LLMs |
| OpenAI GPT-5.6 Terra | 17 | LLMs |
| OpenAI GPT-4.1 | 18 | LLMs |
| Alibaba Machine Translation API | 18 | MT engines |
| Cohere Command A Translate | 30 | LLMs |
| Anthropic Claude Haiku 4.5 | 32 | LLMs |
| xAI Grok 4.20 (reasoning) | 33 | LLMs |
| Swiss AI Apertus-8B-2509 | 34 | Open-weight LLMs |
| OpenAI GPT-5.5 | 39 | LLMs |
| MiniMax M2.5 | 39 | LLMs |
| Mistral Medium 3.1 | 41 | LLMs |
| OpenAI GPT-5.6 Sol | 41 | LLMs |
| Anthropic Claude Sonnet 5 | 43 | LLMs |
| Naver Papago NMT Commercial | 43 | MT engines |
| Youdao Translation Service API | 46 | MT engines |
| iFlyTek MT | 47 | MT engines |
| Widn AI Tower Sugarloaf | 49 | LLM-based MT engines |
| Anthropic Claude Fable 5 | 53 | LLMs |
| Mistral Large 3 | 53 | LLMs |
| Anthropic Claude Opus 4.8 | 54 | LLMs |
| DeepSeek V3.2 | 62 | LLMs |
| AI21 Jamba Large 1.7 | 62 | LLMs |
| Anthropic Claude Sonnet 4.6 | 64 | LLMs |
| AppTek NMT | 85 | MT engines |
| Google Gemini 3.1 Pro Preview | 90 | LLMs |
| Alibaba Qwen3-235B-A22B | 109 | Open-weight LLMs |
| Google Gemma 4 | 157 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 191 | LLMs |
| z.ai GLM-5 | 206 | 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 Korean.
| The two best solutions | Solution | Sentences about people that are gender-neutral | Meets every Advanced requirement | Cost per million characters | Seconds per 1,000 words |
|---|---|---|---|---|---|
| Google Gemini 3.1 Pro Preview | three-step | 100% | ✓ | $577 | 293 |
| Google Gemma 4 | three-step, with the gender-neutrality agent | 100% | ✓ | $1,000 | 596 |
For Korean we check the formal register, with 하십시오체, 합니다체 and honorifics. For gender neutrality we check neutral wording: neutral nouns, omission or rephrasing, with no forced 그 or 그녀 and no default gender. Of the two best solutions, the three-step one (translation + LQA + repair) on Google Gemini 3.1 Pro Preview has 100% of the sentences about people gender-neutral without the gender-neutrality agent. The one on Google Gemma 4 has 100% with the agent.
Leave feedback