The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Hungarian 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 Hungarian. |
| Tone of voice | The formal register (Ön, Önök) 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 Hungarian, 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.
None of the three solutions that meet every Standard requirement leads on errors, cost and time at once. The two-step solution on Google Gemini 3.1 Pro Preview has the fewest total errors, 15.2 per 1,000 words, but takes 204 seconds per 1,000 words. The one-step solution on Anthropic Claude Fable 5 takes 74 seconds per 1,000 words, and the two-step one on OpenAI GPT-5.5 costs the most, $43.30 per million characters. Optimize for selects Google Gemini 3.1 Pro Preview for quality and Anthropic Claude Fable 5 for cost and latency.
A limit of one step leaves only the solution on Anthropic Claude Fable 5 among those that meet every Standard requirement. A cost cap of at most $30.63 per million characters or a time limit below 74 seconds per 1,000 words leaves only solutions that miss general translation quality. The one-step solution on OpenAI GPT-5.6 Luna, for one, costs $0.50 per million characters and takes 18 seconds per 1,000 words, with 24.7 total errors per 1,000 words.
Of the three solutions that meet every Standard requirement, the Advanced requirements leave the two-step solutions on Google Gemini 3.1 Pro Preview and OpenAI GPT-5.5, both without the gender-neutrality agent. Meeting them costs at least $33.03 per million characters and takes at least 146 seconds per 1,000 words, against $30.63 and 74 seconds under the Standard requirements. The two-step solution on Google Gemini 3.1 Pro Preview has the fewest total errors under both sets of requirements, 15.2 per 1,000 words.
Google Gemini 3.1 Pro Preview
two-step
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Google Gemini 3.1 Pro Preview
two-step
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Google Gemini 3.1 Pro Preview
two-step
Google Gemini 3.1 Pro Preview
two-step
OpenAI GPT-5.5
two-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 1: simple solution + backup | ★ ☆☆ | one-step | Anthropic Claude Fable 5 | 16.2 | $30.64 | 74 |
| Standard | Tier 1: simple solution + backup | ★ ☆☆ | one-step | Anthropic Claude Fable 5 | 16.2 | $30.64 | 74 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | two-step | Google Gemini 3.1 Pro Preview | 15.2 | $33.04 | 204 |
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.
None of the three solutions that meet the Basic requirements drop out under the Standard ones, so both levels take one step at the simplest. The Advanced requirements remove the one-step solution and take two steps, without the gender-neutrality agent. Under the Basic and the Standard requirements, Optimize for selects Google Gemini 3.1 Pro Preview for quality and Anthropic Claude Fable 5 for cost and latency. Under the Advanced ones, it selects Google Gemini 3.1 Pro Preview for quality and cost and OpenAI GPT-5.5 for latency.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| OpenAI GPT-5.6 Luna | 0.5 | 23.4 | Misses a Standard requirement |
| DeepSeek V3.2 | 1.07 | 41.2 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 3.44 | 52.4 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.71 | 25.4 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.96 | 23.4 | Misses a Standard requirement |
| Widn AI Tower Sugarloaf | 10 | 78.4 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 12.52 | 21.4 | Misses a Standard requirement |
| Anthropic Claude Opus 4.8 | 15.55 | 18.1 | Misses a Standard requirement |
| Google Translation LLM | 21.41 | 68.2 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 34.3 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 35.9 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 30.64 | 16.2 | Meets every Standard requirement |
Fewer errors come with a higher price along four one-step LLM solutions: 23.4 weighted errors per 1,000 words on OpenAI GPT-5.6 Luna at $0.50 per million characters, 21.4 on OpenAI GPT-5.5 at $12.52, 18.1 on Anthropic Claude Opus 4.8 at $15.55 and 16.2 on Anthropic Claude Fable 5 at $30.64. Each of the last three has fewer weighted errors than every cheaper one-step solution. The one-step solution on Google Gemini 3.1 Pro Preview has 23.4 weighted errors per 1,000 words too, at $4.96 per million characters.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 29.8 |
| Tone of voice | 0.4 |
| Tag handling | 0.3 |
| Gender neutrality | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
Translations from Anthropic Claude Opus 4.8 and OpenAI GPT-5.6 Luna keep 99.2% of the tags, and those from OpenAI GPT-4.1 99.6%, so all three miss tag handling by a small margin. None of the four shortlisted solutions on LLM-based MT engines meets formatting.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Hungarian | Class |
|---|---|---|
| Widn AI Tower Sugarloaf | 78.4 | LLM-based MT engines |
| Google Translation LLM | 68.2 | LLM-based MT engines |
| OpenAI GPT-4.1 | 52.4 | LLMs |
| DeepSeek V3.2 | 41.2 | LLMs |
| DeepL Next Gen | 35.9 | LLM-based MT engines |
| Translated Lara 2 | 34.3 | LLM-based MT engines |
| Google Gemini 3.5 Flash | 25.4 | LLMs |
| Google Gemini 3.1 Pro Preview | 23.4 | LLMs |
| OpenAI GPT-5.6 Luna | 23.4 | LLMs |
| OpenAI GPT-5.5 | 21.4 | LLMs |
| Anthropic Claude Opus 4.8 | 18.1 | LLMs |
| Anthropic Claude Fable 5 | 16.2 | LLMs |
No translation into Hungarian: SpeakLeash Bielik-11B-v3.0-Instruct (wrong language), Widn AI Tower Anthill (cut off part-way), Widn AI Tower Vesuvius (cut off part-way).
The model types overlap on errors. Translations from DeepSeek V3.2 and OpenAI GPT-4.1, both LLMs, have 41.2 and 52.4 weighted errors per 1,000 words, against 34.3 and 35.9 from Translated Lara 2 and DeepL Next Gen, both LLM-based MT engines. All eight shortlisted LLMs support Hungarian by our measured criteria. From three models we got no usable translation: Widn AI Tower Anthill and Widn AI Tower Vesuvius, with translations cut off part-way, and SpeakLeash Bielik-11B-v3.0-Instruct, with text in the wrong language.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Anthropic Claude Sonnet 5 | 0.824 | 9 | 1.60 | 16 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.822 | 10 | 1.57 | 13 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.821 | 11 | 1.50 | 10 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.821 | 12 | 1.59 | 15 | No | Not shortlisted |
| z.ai GLM-5 | 0.815 | 16 | 1.52 | 12 | No | Not shortlisted |
| Mistral Large 3 | 0.813 | 18 | 1.64 | 17 | No | Not shortlisted |
| Tilde TildeOpen LLM engine | 0.811 | 19 | 1.75 | 20 | No | Not shortlisted |
| Cohere Command A Translate | 0.808 | 20 | 1.59 | 14 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.806 | 21 | 1.84 | 23 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.804 | 22 | 1.87 | 25 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.802 | 23 | 1.86 | 24 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.801 | 24 | 1.78 | 22 | No | Not shortlisted |
| MiniMax M2.5 | 0.796 | 25 | 1.75 | 19 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.794 | 26 | 1.90 | 26 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.787 | 27 | 2.44 | 33 | No | Not shortlisted |
| Google Gemma 4 | 0.785 | 28 | 2.12 | 29 | No | Not shortlisted |
| Microsoft Translator | 0.781 | 29 | 1.96 | 27 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.780 | 30 | 2.13 | 30 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.778 | 31 | 2.01 | 28 | No | Not shortlisted |
| NiuTrans | 0.774 | 32 | 2.49 | 35 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.773 | 33 | 2.16 | 31 | No | Not shortlisted |
| Pangeanic PangeaMT | 0.759 | 34 | 2.48 | 34 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.754 | 35 | 2.84 | 37 | No | Not shortlisted |
| Supertext LLM | 0.749 | 36 | 2.34 | 32 | No | Not shortlisted |
| Baidu Translate API | 0.742 | 37 | 2.65 | 36 | No | Not shortlisted |
| Amazon Translate | 0.739 | 38 | 3.37 | 38 | No | Not shortlisted |
| Cohere Tiny Aya | 0.673 | 39 | 4.63 | 39 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.642 | 40 | 5.54 | 40 | No | Not shortlisted |
| AppTek NMT | 0.488 | 41 | 9.43 | 41 | No | Not shortlisted |
| DeepL Next Gen | 0.833 | 1 | 1.37 | 1 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.831 | 2 | 1.40 | 4 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Luna | 0.831 | 3 | 1.38 | 2 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.831 | 4 | 1.42 | 5 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.5 | 0.826 | 5 | 1.52 | 11 | Yes | Top 8 by COMET-Kiwi |
| OpenAI GPT-4.1 | 0.826 | 6 | 1.42 | 6 | Yes | Top 8 by both metrics |
| Google Gemini 3.5 Flash | 0.826 | 7 | 1.44 | 9 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Opus 4.8 | 0.825 | 8 | 1.42 | 7 | Yes | Top 8 by both metrics |
| Google Translation LLM | 0.818 | 13 | 1.77 | 21 | Yes | Real-time place |
| DeepSeek V3.2 | 0.818 | 14 | 1.44 | 8 | Yes | Top 8 by MetricX |
| Translated Lara 2 | 0.816 | 15 | 1.70 | 18 | Yes | Real-time place |
| Widn AI Tower Sugarloaf | 0.814 | 17 | 1.40 | 3 | Yes | Top 8 by MetricX |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| DeepL Next Gen | Off-the-shelf NMT | 1 | 0.833 | 1 | 1.37 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 2 | 0.831 | 4 | 1.40 | Top 8 by both metrics |
| OpenAI GPT-5.6 Luna | Adjusted LLM | 3 | 0.831 | 2 | 1.38 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 4 | 0.831 | 5 | 1.42 | Top 8 by both metrics |
| OpenAI GPT-5.5 | Adjusted LLM | 5 | 0.826 | 11 | 1.52 | Top 8 by COMET-Kiwi |
| OpenAI GPT-4.1 | Adjusted LLM | 6 | 0.826 | 6 | 1.42 | Top 8 by both metrics |
| Google Gemini 3.5 Flash | Adjusted LLM | 7 | 0.826 | 9 | 1.44 | Top 8 by COMET-Kiwi |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 8 | 0.825 | 7 | 1.42 | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 9 | 0.824 | 16 | 1.60 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 10 | 0.822 | 13 | 1.57 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 11 | 0.821 | 10 | 1.50 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 12 | 0.821 | 15 | 1.59 | Not shortlisted |
| Google Translation LLM | Adjusted NMT | 13 | 0.818 | 21 | 1.77 | Real-time place |
| DeepSeek V3.2 | Adjusted LLM | 14 | 0.818 | 8 | 1.44 | Top 8 by MetricX |
| Translated Lara 2 | Adjusted NMT | 15 | 0.816 | 18 | 1.70 | Real-time place |
| z.ai GLM-5 | Adjusted LLM | 16 | 0.815 | 12 | 1.52 | Not shortlisted |
| Widn AI Tower Sugarloaf | Adjusted NMT | 17 | 0.814 | 3 | 1.40 | Top 8 by MetricX |
| Mistral Large 3 | Adjusted LLM | 18 | 0.813 | 17 | 1.64 | Not shortlisted |
| Tilde TildeOpen LLM engine | Off-the-shelf NMT | 19 | 0.811 | 20 | 1.75 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 20 | 0.808 | 14 | 1.59 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 21 | 0.806 | 23 | 1.84 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 22 | 0.804 | 25 | 1.87 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 23 | 0.802 | 24 | 1.86 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 24 | 0.801 | 22 | 1.78 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 25 | 0.796 | 19 | 1.75 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 26 | 0.794 | 26 | 1.90 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 27 | 0.787 | 33 | 2.44 | Not shortlisted |
| Google Gemma 4 | Adjusted LLM | 28 | 0.785 | 29 | 2.12 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 29 | 0.781 | 27 | 1.96 | Not shortlisted |
| SYSTRAN PNMT API | Off-the-shelf NMT | 30 | 0.780 | 30 | 2.13 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 31 | 0.778 | 28 | 2.01 | Not shortlisted |
| NiuTrans | Off-the-shelf NMT | 32 | 0.774 | 35 | 2.49 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 33 | 0.773 | 31 | 2.16 | Not shortlisted |
| Pangeanic PangeaMT | Off-the-shelf NMT | 34 | 0.759 | 34 | 2.48 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 35 | 0.754 | 37 | 2.84 | Not shortlisted |
| Supertext LLM | Off-the-shelf NMT | 36 | 0.749 | 32 | 2.34 | Not shortlisted |
| Baidu Translate API | Off-the-shelf NMT | 37 | 0.742 | 36 | 2.65 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 38 | 0.739 | 38 | 3.37 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 39 | 0.673 | 39 | 4.63 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 40 | 0.642 | 40 | 5.54 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 41 | 0.488 | 41 | 9.43 | Not shortlisted |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.827 | — | 1.32 | Same model family as OpenAI GPT-5.6 Luna |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.824 | — | 1.55 | Same model family as OpenAI GPT-5.6 Luna |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.821 | — | 1.61 | 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.
The metrics and Intento LQA order the shortlisted solutions differently. Translations from DeepL Next Gen, first on both metrics, have 35.9 weighted errors per 1,000 words by Intento LQA, against 16.2 from Anthropic Claude Fable 5, second on COMET-Kiwi and fourth on MetricX. Translations from Widn AI Tower Sugarloaf, third on MetricX and 17th on COMET-Kiwi, have 78.4 weighted errors per 1,000 words. Nearest the cut are Anthropic Claude Sonnet 5, rank 9 on COMET-Kiwi, and Anthropic Claude Sonnet 4.6, rank 10 on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | $30.64 |
| Standard | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | $30.64 |
| Advanced | Google Gemini 3.1 Pro Preview, two-step | $33.04 |
| Model | US dollars per million characters | Class |
|---|---|---|
| OpenAI GPT-5.6 Luna | $0.5 | LLMs |
| DeepSeek V3.2 | $1.07 | LLMs |
| OpenAI GPT-4.1 | $3.44 | LLMs |
| Google Gemini 3.5 Flash | $3.71 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.96 | LLMs |
| Widn AI Tower Sugarloaf | $10 | LLM-based MT engines |
| OpenAI GPT-5.5 | $12.52 | LLMs |
| Anthropic Claude Opus 4.8 | $15.55 | LLMs |
| Google Translation LLM | $21.41 | LLM-based MT engines |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $30.64 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Sonnet 5 | 0.479 | LLMs |
| Anthropic Claude Fable 5 | 0.478 | LLMs |
| Anthropic Claude Opus 4.8 | 0.478 | LLMs |
| AI21 Jamba Large 1.7 | 0.448 | LLMs |
| Mistral Medium 3.1 | 0.439 | LLMs |
| Mistral Large 3 | 0.437 | LLMs |
| Widn AI Tower Sugarloaf | 0.409 | LLM-based MT engines |
| Alibaba Qwen3-235B-A22B | 0.403 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 0.403 | LLMs |
| Cohere Command A Translate | 0.402 | LLMs |
| DeepSeek V3.2 | 0.378 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.374 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.371 | LLMs |
| z.ai GLM-5 | 0.354 | LLMs |
| OpenAI GPT-4.1 | 0.335 | LLMs |
| MiniMax M2.5 | 0.334 | LLMs |
| OpenAI GPT-5.5 | 0.333 | LLMs |
| OpenAI GPT-5.6 Sol | 0.333 | LLMs |
| OpenAI GPT-5.6 Terra | 0.333 | LLMs |
| OpenAI GPT-5.6 Luna | 0.332 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.325 | LLMs |
| Google Gemini 3.5 Flash | 0.324 | LLMs |
| Google Gemma 4 | 0.32 | Open-weight LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.292 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.29 | LLMs |
| Cohere Tiny Aya | 0.28 | LLMs |
The tokenizers produce a median 0.36 tokens per character of Hungarian, 1.4 times their count for English. The two dearest one-step LLM solutions, on Anthropic Claude Fable 5 and Anthropic Claude Opus 4.8, cost $30.64 and $15.55 per million characters. Their tokenizers produce 0.478 tokens per character, against 0.324 to 0.378 for the other six shortlisted LLMs. The post-edit step takes the solution on Google Gemini 3.1 Pro Preview from $4.96 to $33.04 per million characters, the lowest cost of meeting every Advanced requirement.
| 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 | 74 |
| Standard | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | 74 |
| Advanced | OpenAI GPT-5.5, two-step | 147 |
| 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 | 0.3 | MT engines |
| Pangeanic PangeaMT | 0.5 | MT engines |
| Google Translation LLM | 0.6 | LLM-based MT engines |
| DeepL Next Gen | 0.8 | LLM-based MT engines |
| Amazon Translate | 1.2 | MT engines |
| Translated Lara 2 | 1.5 | LLM-based MT engines |
| SYSTRAN PNMT API | 2.2 | MT engines |
| OpenAI gpt-oss-120b | 2.6 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 4.3 | MT engines |
| Tilde TildeOpen LLM engine | 9.3 | LLM-based MT engines |
| Baidu Translate API | 10 | MT engines |
| OpenAI GPT-5.6 Luna | 18 | LLMs |
| Supertext LLM | 18 | LLM-based MT engines |
| xAI Grok 4.20 (non-reasoning) | 21 | LLMs |
| OpenAI GPT-4.1 | 25 | LLMs |
| Cohere Tiny Aya | 31 | LLMs |
| OpenAI GPT-5.6 Terra | 31 | LLMs |
| Anthropic Claude Haiku 4.5 | 39 | LLMs |
| MiniMax M2.5 | 40 | LLMs |
| Mistral Medium 3.1 | 42 | LLMs |
| Alibaba Machine Translation API | 43 | MT engines |
| Widn AI Tower Sugarloaf | 51 | LLM-based MT engines |
| NiuTrans | 56 | MT engines |
| Mistral Large 3 | 58 | LLMs |
| AppTek NMT | 58 | MT engines |
| DeepSeek V3.2 | 59 | LLMs |
| Google Gemini 3.5 Flash | 61 | LLMs |
| OpenAI GPT-5.6 Sol | 61 | LLMs |
| Cohere Command A Translate | 63 | LLMs |
| AI21 Jamba Large 1.7 | 63 | LLMs |
| OpenAI GPT-5.5 | 72 | LLMs |
| Anthropic Claude Fable 5 | 74 | LLMs |
| Anthropic Claude Sonnet 5 | 77 | LLMs |
| xAI Grok 4.20 (reasoning) | 78 | LLMs |
| Anthropic Claude Sonnet 4.6 | 78 | LLMs |
| Google Gemini 3.1 Pro Preview | 89 | LLMs |
| Anthropic Claude Opus 4.8 | 95 | LLMs |
| Swiss AI Apertus-8B-2509 | 101 | Open-weight LLMs |
| Google Gemma 4 | 166 | Open-weight LLMs |
| Alibaba Qwen3-235B-A22B | 170 | Open-weight LLMs |
| z.ai GLM-5 | 251 | LLMs |
| MoonshotAI Kimi K2.5 | 255 | 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 Hungarian.
The three real-time solutions on the shortlist, on Google Translation LLM, DeepL Next Gen and Translated Lara 2, have 68.2, 35.9 and 34.3 weighted errors per 1,000 words. Nine of the ten real-time models are MT engines or LLM-based MT engines, and the other is the open-weight OpenAI gpt-oss-120b, which did not make the shortlist. The fastest shortlisted one-step LLM solution, on OpenAI GPT-5.6 Luna, takes 18 seconds per 1,000 words. The two solutions that meet every Advanced requirement take 147 and 204 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 |
|---|---|---|---|---|---|
| Google Gemini 3.1 Pro Preview | two-step | 100% | ✓ | $33.04 | 204 |
| OpenAI GPT-5.5 | two-step | 100% | ✓ | $43.30 | 147 |
For tone of voice we check the formal register, with Ön and Önök as the form of address. Translations from every shortlisted LLM have 100% of the sentences in the formal register, from DeepL Next Gen 76.6% and from Google Translation LLM 89.4%. For gender neutrality we check that a translation does not add a gender the source does not specify, with no default gender. Hungarian has no grammatical gender for persons. The two best solutions have 100% of the sentences about people gender-neutral without the gender-neutrality agent.
Each pair of automatic metrics agrees above its threshold in Hungarian, and neither sits on the border. The three LQA judges agree on a median of 95.2% of the sentences of the shortlisted solutions. So the quality results on this page rest on measurement we can rely on.
Leave feedback