The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Romanian 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 (dumneavoastră) 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 Romanian, 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.
Both solutions that meet the Standard requirements trade cost against errors and time. The two-step solution (translation + post-edit) on OpenAI GPT-5.6 Sol has 25.9 total errors per 1,000 words, costs $19.29 per million characters and takes 95 seconds per 1,000 words. The one-step solution on Anthropic Claude Fable 5 has 25.4 total errors per 1,000 words, costs $29.18 per million characters and takes 60 seconds per 1,000 words. Optimize for selects Anthropic Claude Fable 5 for quality and latency, and OpenAI GPT-5.6 Sol for cost.
A cost cap below $19.29 per million characters excludes every solution that meets the Basic or Standard requirements. A time limit of 59 seconds per 1,000 words or less excludes both solutions that meet the Standard requirements. A limit of one step leaves only the solution on Anthropic Claude Fable 5 under the Standard requirements. Under the Basic requirements, the same limit also leaves DeepL Next Gen, the one real-time (<2 sec) solution that meets them.
For the Advanced requirements, the three-step solution (translation + LQA + repair) on Anthropic Claude Fable 5 with the gender-neutrality agent costs $1,029 per million characters and takes 216 seconds per 1,000 words, against $29.18 and 60 seconds in one step. Total errors fall from 25.4 to 7.9 per 1,000 words. With the agent, the two-step solution on OpenAI GPT-5.6 Sol costs $19.45 per million characters against $19.29, and its total errors fall from 25.9 to 16.5 per 1,000 words.
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
two-step
DeepL Next Gen
one-step NMT adjusted for Standard requirements
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
two-step
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
Anthropic Claude Fable 5
three-step with the gender-neutrality agent
Anthropic Claude Fable 5
three-step with the gender-neutrality agent
Anthropic Claude Fable 5
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 | DeepL Next Gen | 40.8 | $27.50 | 1.0 |
| Standard | Tier 1: simple solution + backup | ★ ☆ | one-step | Anthropic Claude Fable 5 | 25.4 | $29.18 | 60 |
| Advanced | Tier 3: one solution | ☆ | with the gender-neutrality agent | Anthropic Claude Fable 5 | 7.9 | $1,029 | 216 |
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.
The Standard requirements remove the one-step solution on DeepL Next Gen, which meets the Basic ones, and keep the one-step solution on Anthropic Claude Fable 5 and the two-step solution (translation + post-edit) on OpenAI GPT-5.6 Sol. Under the Basic and Standard requirements, Optimize for selects the one on Anthropic Claude Fable 5 for quality and the one on OpenAI GPT-5.6 Sol for cost. For latency it selects the one on DeepL Next Gen under the Basic requirements and the one on Anthropic Claude Fable 5 under the Standard ones.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Mistral Large 3 | 0.79 | 44.5 | Misses a Standard requirement |
| z.ai GLM-5 | 1.92 | 27.7 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 3.15 | 36.3 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.57 | 26 | Misses a Standard requirement |
| Anthropic Claude Sonnet 4.6 | 6.61 | 29.3 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 7.73 | 20.4 | Misses a Standard requirement |
| Widn AI Tower Sugarloaf | 10 | 32.3 | Misses a Standard requirement |
| Microsoft Translator | 10 | 57.7 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 32 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 17.5 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 29.18 | 16.2 | Meets every Standard requirement |
Apart from the two dearest one-step solutions, a higher price does not always come with fewer errors. The one on Translated Lara 2 costs $24.99 per million characters and has 32.0 weighted errors per 1,000 words, against $7.73 and 20.4 on OpenAI GPT-5.6 Sol and $4.57 and 26.0 on Google Gemini 3.1 Pro Preview. The cheapest, on Mistral Large 3, costs $0.79 per million characters and has 44.5 weighted errors per 1,000 words.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 28.5 |
| Gender neutrality | 10.9 |
| Tone of voice | 1.1 |
| Tag handling | 0.3 |
| Terminology | 0.2 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
Translations from the three LLM-based MT engines come close on terminology, with 97.1% to 98.8% of the glossary terms right, and six of the eight LLM solutions meet it. On tag handling, translations from Cohere Command A Translate have 99.6% of the tags correct. The two-step solution (translation + post-edit) on Google Gemini 3.1 Pro Preview has 19.8 weighted errors per 1,000 words against 26.0 in one step, but 4 major errors, more than general translation quality allows.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Romanian | Class |
|---|---|---|
| Microsoft Translator | 57.7 | MT engines |
| Mistral Large 3 | 44.5 | LLMs |
| OpenAI GPT-4.1 | 36.3 | LLMs |
| Widn AI Tower Sugarloaf | 32.3 | LLM-based MT engines |
| Translated Lara 2 | 32 | LLM-based MT engines |
| Anthropic Claude Sonnet 4.6 | 29.3 | LLMs |
| z.ai GLM-5 | 27.7 | LLMs |
| Google Gemini 3.1 Pro Preview | 26 | LLMs |
| Cohere Command A Translate | 22.1 | LLMs |
| OpenAI GPT-5.6 Sol | 20.4 | LLMs |
| DeepL Next Gen | 17.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | 16.2 | LLMs |
No translation into Romanian: SpeakLeash Bielik-11B-v3.0-Instruct (wrong language), Widn AI Tower Vesuvius (cut off part-way).
The eight shortlisted LLM solutions have 16.2 to 44.5 weighted errors per 1,000 words, the three LLM-based MT engines 17.5 to 32.3, and Microsoft Translator, the only shortlisted MT engine, 57.7. Translations from DeepL Next Gen, the LLM-based MT engine with the fewest, have 1.08 times as many weighted errors as those from Anthropic Claude Fable 5, the LLM with the fewest. All eight LLMs support Romanian by our measured criteria. Of their one-step solutions, only the one on Anthropic Claude Fable 5 meets general translation quality.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Anthropic Claude Sonnet 5 | 0.893 | 9 | 1.73 | 17 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.892 | 11 | 1.70 | 11 | No | Not shortlisted |
| DeepSeek V3.2 | 0.892 | 12 | 1.74 | 18 | No | Not shortlisted |
| Anthropic Claude Opus 4.8 | 0.892 | 13 | 1.73 | 15 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.891 | 14 | 1.72 | 14 | No | Not shortlisted |
| Google Gemini 3.5 Flash | 0.891 | 16 | 1.71 | 12 | No | Not shortlisted |
| Google Gemma 4 | 0.888 | 19 | 1.70 | 10 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.888 | 20 | 1.79 | 21 | No | Not shortlisted |
| OpenAI GPT-5.5 | 0.887 | 21 | 1.76 | 19 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.887 | 22 | 1.72 | 13 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.886 | 23 | 1.86 | 25 | No | Not shortlisted |
| Pangeanic PangeaMT | 0.885 | 24 | 1.81 | 23 | No | Not shortlisted |
| Google Translation LLM | 0.884 | 25 | 2.01 | 27 | No | Not shortlisted |
| Tilde TildeOpen LLM engine | 0.883 | 26 | 2.03 | 29 | No | Not shortlisted |
| Cohere Tiny Aya | 0.881 | 27 | 2.02 | 28 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.878 | 28 | 1.81 | 22 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.875 | 29 | 1.97 | 26 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.869 | 30 | 2.12 | 31 | No | Not shortlisted |
| MiniMax M2.5 | 0.868 | 31 | 2.25 | 32 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.861 | 32 | 2.45 | 35 | No | Not shortlisted |
| Amazon Translate | 0.857 | 33 | 2.40 | 34 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.857 | 34 | 2.11 | 30 | No | Not shortlisted |
| Baidu Translate API | 0.852 | 35 | 2.35 | 33 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.851 | 36 | 2.70 | 38 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.847 | 37 | 2.53 | 36 | No | Not shortlisted |
| Oracle Cloud Language Service | 0.847 | 38 | 2.59 | 37 | No | Not shortlisted |
| AppTek NMT | 0.807 | 39 | 3.34 | 39 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.774 | 40 | 3.66 | 40 | No | Not shortlisted |
| Widn AI Tower Anthill | 0.493 | 41 | 11.82 | 41 | No | Not shortlisted |
| DeepL Next Gen | 0.902 | 1 | 1.50 | 1 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | 0.900 | 2 | 1.53 | 2 | Yes | Top 8 by both metrics |
| OpenAI GPT-4.1 | 0.898 | 3 | 1.69 | 8 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.897 | 4 | 1.65 | 5 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.897 | 5 | 1.70 | 9 | Yes | Top 8 by COMET-Kiwi |
| Cohere Command A Translate | 0.896 | 6 | 1.60 | 4 | Yes | Top 8 by both metrics |
| Widn AI Tower Sugarloaf | 0.895 | 7 | 1.58 | 3 | Yes | Top 8 by both metrics |
| z.ai GLM-5 | 0.893 | 8 | 1.77 | 20 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 4.6 | 0.893 | 10 | 1.68 | 7 | Yes | Top 8 by MetricX |
| Mistral Large 3 | 0.891 | 15 | 1.66 | 6 | Yes | Top 8 by MetricX |
| Translated Lara 2 | 0.889 | 17 | 1.82 | 24 | Yes | Real-time place |
| Microsoft Translator | 0.889 | 18 | 1.73 | 16 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| DeepL Next Gen | Adjusted NMT | 1 | 0.902 | 1 | 1.50 | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | Adjusted LLM | 2 | 0.900 | 2 | 1.53 | Top 8 by both metrics |
| OpenAI GPT-4.1 | Adjusted LLM | 3 | 0.898 | 8 | 1.69 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 4 | 0.897 | 5 | 1.65 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 5 | 0.897 | 9 | 1.70 | Top 8 by COMET-Kiwi |
| Cohere Command A Translate | Adjusted LLM | 6 | 0.896 | 4 | 1.60 | Top 8 by both metrics |
| Widn AI Tower Sugarloaf | Adjusted NMT | 7 | 0.895 | 3 | 1.58 | Top 8 by both metrics |
| z.ai GLM-5 | Adjusted LLM | 8 | 0.893 | 20 | 1.77 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 9 | 0.893 | 17 | 1.73 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 10 | 0.893 | 7 | 1.68 | Top 8 by MetricX |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 11 | 0.892 | 11 | 1.70 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 12 | 0.892 | 18 | 1.74 | Not shortlisted |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 13 | 0.892 | 15 | 1.73 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 14 | 0.891 | 14 | 1.72 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 15 | 0.891 | 6 | 1.66 | Top 8 by MetricX |
| Google Gemini 3.5 Flash | Adjusted LLM | 16 | 0.891 | 12 | 1.71 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 17 | 0.889 | 24 | 1.82 | Real-time place |
| Microsoft Translator | Off-the-shelf NMT | 18 | 0.889 | 16 | 1.73 | Real-time place |
| Google Gemma 4 | Adjusted LLM | 19 | 0.888 | 10 | 1.70 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Off-the-shelf LLM | 20 | 0.888 | 21 | 1.79 | Not shortlisted |
| OpenAI GPT-5.5 | Adjusted LLM | 21 | 0.887 | 19 | 1.76 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 22 | 0.887 | 13 | 1.72 | Not shortlisted |
| OpenAI gpt-oss-120b | Off-the-shelf LLM | 23 | 0.886 | 25 | 1.86 | Not shortlisted |
| Pangeanic PangeaMT | Off-the-shelf NMT | 24 | 0.885 | 23 | 1.81 | Not shortlisted |
| Google Translation LLM | Off-the-shelf NMT | 25 | 0.884 | 27 | 2.01 | Not shortlisted |
| Tilde TildeOpen LLM engine | Adjusted NMT | 26 | 0.883 | 29 | 2.03 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 27 | 0.881 | 28 | 2.02 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 28 | 0.878 | 22 | 1.81 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 29 | 0.875 | 26 | 1.97 | Not shortlisted |
| ModernMT Enterprise Edition | Adjusted NMT | 30 | 0.869 | 31 | 2.12 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 31 | 0.868 | 32 | 2.25 | Not shortlisted |
| SYSTRAN PNMT API | Adjusted NMT | 32 | 0.861 | 35 | 2.45 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 33 | 0.857 | 34 | 2.40 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 34 | 0.857 | 30 | 2.11 | Not shortlisted |
| Baidu Translate API | Off-the-shelf NMT | 35 | 0.852 | 33 | 2.35 | Not shortlisted |
| Google Cloud Advanced Translation API | Adjusted NMT | 36 | 0.851 | 38 | 2.70 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 37 | 0.847 | 36 | 2.53 | Not shortlisted |
| Oracle Cloud Language Service | Off-the-shelf NMT | 38 | 0.847 | 37 | 2.59 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 39 | 0.807 | 39 | 3.34 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 40 | 0.774 | 40 | 3.66 | Not shortlisted |
| Widn AI Tower Anthill | Adjusted NMT | 41 | 0.493 | 41 | 11.82 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.898 | — | 1.75 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.898 | — | 1.58 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.883 | — | 1.85 | 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. Anthropic Claude Fable 5 is on the list through COMET-Kiwi alone, 5th there and 9th on MetricX. Its one-step solution is the only one-step solution that meets every Standard requirement. Translations from Widn AI Tower Sugarloaf, 3rd on MetricX, have 32.3 weighted errors per 1,000 words, and from OpenAI GPT-4.1, 3rd on COMET-Kiwi, 36.3. The models nearest the cut are Anthropic Claude Sonnet 5, 9th on COMET-Kiwi, and Google Gemma 4, 10th on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.6 Sol, two-step | $19.29 |
| Standard | OpenAI GPT-5.6 Sol, two-step | $19.29 |
| Advanced | Anthropic Claude Fable 5, three-step with the gender-neutrality agent | $1,029 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Mistral Large 3 | $0.79 | LLMs |
| z.ai GLM-5 | $1.92 | LLMs |
| OpenAI GPT-4.1 | $3.15 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.57 | LLMs |
| Anthropic Claude Sonnet 4.6 | $6.61 | LLMs |
| OpenAI GPT-5.6 Sol | $7.73 | LLMs |
| Widn AI Tower Sugarloaf | $10 | LLM-based MT engines |
| Microsoft Translator | $10 | MT engines |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $29.18 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Fable 5 | 0.435 | LLMs |
| Anthropic Claude Sonnet 5 | 0.435 | LLMs |
| Anthropic Claude Opus 4.8 | 0.434 | LLMs |
| AI21 Jamba Large 1.7 | 0.384 | LLMs |
| Mistral Medium 3.1 | 0.369 | LLMs |
| Mistral Large 3 | 0.367 | LLMs |
| Widn AI Tower Anthill | 0.336 | LLM-based MT engines |
| Alibaba Qwen3-235B-A22B | 0.332 | Open-weight LLMs |
| Anthropic Claude Sonnet 4.6 | 0.332 | LLMs |
| Widn AI Tower Sugarloaf | 0.332 | LLM-based MT engines |
| DeepSeek V3.2 | 0.331 | LLMs |
| MoonshotAI Kimi K2.5 | 0.331 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.329 | LLMs |
| z.ai GLM-5 | 0.306 | LLMs |
| OpenAI GPT-4.1 | 0.291 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.29 | LLMs |
| Google Gemini 3.5 Flash | 0.29 | LLMs |
| MiniMax M2.5 | 0.289 | LLMs |
| OpenAI GPT-5.5 | 0.289 | LLMs |
| Google Gemma 4 | 0.288 | Open-weight LLMs |
| OpenAI GPT-5.6 Sol | 0.288 | LLMs |
| OpenAI GPT-5.6 Terra | 0.288 | LLMs |
| OpenAI GPT-5.6 Luna | 0.287 | LLMs |
| Cohere Command A Translate | 0.274 | LLMs |
| Cohere Tiny Aya | 0.273 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.271 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.271 | LLMs |
Tokenizers produce a median 0.31 tokens per character of Romanian, 1.2 times their count for English. The tokenizer of Mistral Large 3 produces 0.367 tokens per character and that of OpenAI GPT-4.1 0.291, but their one-step solutions cost $0.79 and $3.15 per million characters. The steps raise the cost. The two-step solutions (translation + post-edit) on OpenAI GPT-5.6 Sol and Google Gemini 3.1 Pro Preview cost $19.29 and $29.29 per million characters, against $7.73 and $4.57 in one step.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | DeepL Next Gen, one-step NMT adjusted for Standard requirements | 0.7 |
| Standard | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | 60 |
| Advanced | Anthropic Claude Fable 5, three-step with the gender-neutrality agent | 216 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.3 | MT engines |
| Google Cloud Advanced Translation API | 0.3 | MT engines |
| Google Translation LLM | 0.4 | LLM-based MT engines |
| Pangeanic PangeaMT | 0.6 | MT engines |
| DeepL Next Gen | 0.7 | LLM-based MT engines |
| ModernMT Enterprise Edition | 0.9 | MT engines |
| SYSTRAN PNMT API | 1.1 | MT engines |
| Amazon Translate | 1.3 | MT engines |
| Translated Lara 2 | 1.4 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 1.8 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 3 | MT engines |
| Oracle Cloud Language Service | 4.5 | MT engines |
| Tilde TildeOpen LLM engine | 7.1 | LLM-based MT engines |
| Widn AI Tower Anthill | 7.9 | LLM-based MT engines |
| Baidu Translate API | 11 | MT engines |
| OpenAI GPT-5.6 Luna | 12 | LLMs |
| Cohere Tiny Aya | 20 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 21 | LLMs |
| OpenAI GPT-5.6 Terra | 22 | LLMs |
| Alibaba Machine Translation API | 25 | MT engines |
| DeepSeek V3.2 | 32 | LLMs |
| Anthropic Claude Haiku 4.5 | 32 | LLMs |
| OpenAI GPT-4.1 | 34 | LLMs |
| Cohere Command A Translate | 39 | LLMs |
| MiniMax M2.5 | 40 | LLMs |
| Mistral Medium 3.1 | 41 | LLMs |
| OpenAI GPT-5.6 Sol | 43 | LLMs |
| Swiss AI Apertus-8B-2509 | 43 | Open-weight LLMs |
| Widn AI Tower Sugarloaf | 49 | LLM-based MT engines |
| Mistral Large 3 | 52 | LLMs |
| Anthropic Claude Sonnet 5 | 53 | LLMs |
| OpenAI GPT-5.5 | 56 | LLMs |
| AI21 Jamba Large 1.7 | 58 | LLMs |
| Anthropic Claude Sonnet 4.6 | 59 | LLMs |
| Anthropic Claude Fable 5 | 60 | LLMs |
| Anthropic Claude Opus 4.8 | 69 | LLMs |
| Google Gemini 3.5 Flash | 70 | LLMs |
| xAI Grok 4.20 (reasoning) | 78 | LLMs |
| Google Gemini 3.1 Pro Preview | 92 | LLMs |
| AppTek NMT | 93 | MT engines |
| Alibaba Qwen3-235B-A22B | 122 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 156 | LLMs |
| z.ai GLM-5 | 172 | LLMs |
| Google Gemma 4 | 205 | Open-weight LLMs |
Real-time models are those the shortlist counts as real-time, under 2 seconds per text across the languages we timed. The chart shows the time of each model into Romanian.
Nine of the 10 real-time models in Romanian are MT engines or LLM-based MT engines, and the other is the open-weight OpenAI gpt-oss-120b, which is not on the shortlist. The one-step LLM solutions on the shortlist take from 34 seconds per 1,000 words, on OpenAI GPT-4.1, to 172, on z.ai GLM-5. The fastest solution that meets every Basic requirement takes 0.7 seconds per 1,000 words, against 60 under the Standard requirements and 216 under the Advanced ones.
| 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 |
|---|---|---|---|---|---|
| Anthropic Claude Fable 5 | three-step, with the gender-neutrality agent | 100% | ✓ | $1,029 | 216 |
| OpenAI GPT-5.6 Sol | two-step, with the gender-neutrality agent | 99.1% | ✗ | $19.45 | 86 |
For tone of voice, we check the formal register, with dumneavoastră as the form of address. For gender neutrality, we check neutral nouns and rephrasing, with paired forms only when unavoidable and never a default gender. The one-step solutions have a median 10.9 gender-neutrality errors per 1,000 words. With the gender-neutrality agent, 100% of the sentences about people in the translations from the three-step solution (translation + LQA + repair) on Anthropic Claude Fable 5 are gender-neutral, at $1,029 per million characters.
The agreement of both pairs of automatic metrics is above the threshold in Romanian, and neither pair is borderline. The three LQA judges agree on a median 95.8% of the sentences of the 12 shortlisted solutions. The results on this page rest on measurement we can rely on.
Leave feedback