The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Serbian 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 (Vi) 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 Serbian, with a column for each requirement you select. Under Optimize for, choose which of them comes first: the one with the fewest total errors, the lowest cost or the shortest time. Click a column heading to sort by it.
The shortlisted solutions on 13 models, with their Intento LQA results. Total errors per 1,000 words add the weighted general translation quality errors (minor 1, major 5, critical 10) to the missed glossary terms, sentences in the wrong register, tag errors and sentences that fail gender neutrality, each counted once. Formatting counts as met or missed and does not add to the total errors. The models that did not make the shortlist have automatic scores only: the shortlist. Gender neutrality: the Advanced requirements count the two best solutions of the language, with the gender-neutrality agent where they need it. The results of the other solutions are shown for information.
Three solutions meet every Standard requirement in Serbian with 24.4 to 26.0 total errors per 1,000 words, so they differ mainly in cost and time. Per million characters, the two-step solution (translation + post-edit) on OpenAI GPT-5.6 Sol costs $16.90, the one-step solution on Anthropic Claude Fable 5 $29.08 and the two-step solution on OpenAI GPT-5.5 $29.39. Per 1,000 words, they take 97, 62 and 112 seconds. Optimize for selects Anthropic Claude Fable 5 for quality and latency, and OpenAI GPT-5.6 Sol for cost.
A limit of one step leaves one solution that meets every Standard requirement, the one on Anthropic Claude Fable 5. A cost cap of $16.89 per million characters or less leaves only solutions that meet every Basic requirement, the one-step solutions on OpenAI GPT-5.6 Sol at $7.50 and OpenAI GPT-5.5 at $10.78. A time limit below 62 seconds per 1,000 words leaves only one solution that meets every Basic requirement, the one-step solution on OpenAI GPT-5.6 Sol at 60 seconds. It misses terminology, one of the Standard requirements.
With the gender-neutrality agent, the two-step solution on OpenAI GPT-5.6 Sol costs $18.58 per million characters instead of $16.90 and takes 103 seconds per 1,000 words instead of 97. Its total errors fall from 25.0 to 19.5 per 1,000 words. The three-step solution (translation + LQA + repair) on Anthropic Claude Fable 5 with the agent, the only one that meets every Advanced requirement, has 11.9 total errors per 1,000 words.
Anthropic Claude Fable 5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
one-step LLM 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 | OpenAI GPT-5.6 Sol | 25.5 | $7.50 | 60 |
| Standard | Tier 1: simple solution + backup | ★ ☆☆ | one-step | Anthropic Claude Fable 5 | 24.4 | $29.08 | 62 |
| Advanced | Tier 3: one solution | ☆ | with the gender-neutrality agent | Anthropic Claude Fable 5 | 11.9 | $1,034 | 224 |
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.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.14 | 37.2 | Misses a Standard requirement |
| Mistral Large 3 | 0.8 | 33.9 | Misses a Standard requirement |
| Anthropic Claude Haiku 4.5 | 2.13 | 98.9 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.39 | 30 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.53 | 33 | Misses a Standard requirement |
| Anthropic Claude Sonnet 5 | 5.74 | 37.9 | Misses a Standard requirement |
| Anthropic Claude Sonnet 4.6 | 6.45 | 41.5 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 7.5 | 18.1 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 10.78 | 19.1 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 33 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 29.08 | 17.8 | Meets every Standard requirement |
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 33.9 |
| Gender neutrality | 7.2 |
| Terminology | 0.5 |
| Tone of voice | 0 |
| Tag handling | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
The one-step solutions that miss tag handling do so by small margins, with 95.7% to 99.2% of the tags right. The ones that miss general translation quality have 30.0 to 98.9 weighted errors per 1,000 words, well above the bar. The two-step solution (translation + post-edit) on Google Gemini 3.1 Pro Preview meets every Standard requirement except general translation quality, with 23.4 weighted errors per 1,000 words.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Serbian | Class |
|---|---|---|
| Anthropic Claude Haiku 4.5 | 98.9 | LLMs |
| ModernMT Enterprise Edition | 94.6 | MT engines |
| Cohere Command A Translate | 80.8 | LLMs |
| Anthropic Claude Sonnet 4.6 | 41.5 | LLMs |
| Anthropic Claude Sonnet 5 | 37.9 | LLMs |
| Google Gemma 4 | 37.2 | Open-weight LLMs |
| Mistral Large 3 | 33.9 | LLMs |
| Google Gemini 3.1 Pro Preview | 33 | LLMs |
| Translated Lara 2 | 33 | LLM-based MT engines |
| Google Gemini 3.5 Flash | 30 | LLMs |
| OpenAI GPT-5.5 | 19.1 | LLMs |
| OpenAI GPT-5.6 Sol | 18.1 | LLMs |
| Anthropic Claude Fable 5 | 17.8 | LLMs |
No translation into Serbian: DeepL Next Gen (wrong language, wrong script), Google Cloud Advanced Translation API (wrong language, wrong script), Google Translation LLM (wrong language, wrong script), SpeakLeash Bielik-11B-v3.0-Instruct (wrong language), Tilde TildeOpen LLM engine (wrong language, wrong script), Youdao Translation Service API (cut off part-way, wrong language, wrong script).
The three one-step solutions that meet general translation quality in Serbian are all on LLMs, with 17.8 to 19.1 weighted errors per 1,000 words, and the solutions on the other eight shortlisted LLMs have 30.0 to 98.9. Translations from Translated Lara 2, an LLM-based MT engine, have 33.0 weighted errors per 1,000 words, and from ModernMT Enterprise Edition, an MT engine, 94.6. From six models, among them DeepL Next Gen and Google Translation LLM, we got text in the wrong language or script.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| OpenAI GPT-4.1 | 0.823 | 10 | 1.52 | 17 | No | Not shortlisted |
| z.ai GLM-5 | 0.823 | 11 | 1.48 | 13 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.819 | 13 | 1.55 | 20 | No | Not shortlisted |
| Anthropic Claude Opus 4.8 | 0.818 | 14 | 1.52 | 18 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.818 | 15 | 1.48 | 12 | No | Not shortlisted |
| DeepSeek V3.2 | 0.818 | 16 | 1.48 | 14 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.816 | 17 | 1.55 | 19 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.815 | 18 | 1.55 | 21 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.813 | 19 | 1.59 | 23 | No | Not shortlisted |
| MiniMax M2.5 | 0.812 | 21 | 1.57 | 22 | No | Not shortlisted |
| Microsoft Translator | 0.800 | 24 | 1.65 | 24 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.793 | 25 | 1.79 | 26 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.786 | 26 | 1.77 | 25 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.782 | 27 | 1.82 | 28 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.780 | 28 | 1.81 | 27 | No | Not shortlisted |
| Cohere Tiny Aya | 0.763 | 29 | 2.16 | 29 | No | Not shortlisted |
| Supertext LLM | 0.753 | 30 | 2.77 | 30 | No | Not shortlisted |
| AppTek NMT | 0.580 | 31 | 6.10 | 31 | No | Not shortlisted |
| OpenAI GPT-5.6 Sol | 0.828 | 1 | 1.39 | 4 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.5 | 0.827 | 2 | 1.47 | 10 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 4.6 | 0.826 | 3 | 1.51 | 16 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | 0.826 | 4 | 1.42 | 7 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.826 | 5 | 1.41 | 6 | Yes | Top 8 by both metrics |
| Google Gemini 3.5 Flash | 0.825 | 6 | 1.38 | 3 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.825 | 7 | 1.46 | 9 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Fable 5 | 0.824 | 8 | 1.48 | 11 | Yes | Top 8 by COMET-Kiwi |
| Translated Lara 2 | 0.823 | 9 | 1.49 | 15 | Yes | Real-time place |
| Mistral Large 3 | 0.822 | 12 | 1.33 | 1 | Yes | Top 8 by MetricX |
| Cohere Command A Translate | 0.813 | 20 | 1.37 | 2 | Yes | Top 8 by MetricX |
| Anthropic Claude Haiku 4.5 | 0.806 | 22 | 1.40 | 5 | Yes | Top 8 by MetricX |
| ModernMT Enterprise Edition | 0.803 | 23 | 1.44 | 8 | Yes | Top 8 by MetricX |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| OpenAI GPT-5.6 Sol | Adjusted LLM | 1 | 0.828 | 4 | 1.39 | Top 8 by both metrics |
| OpenAI GPT-5.5 | Adjusted LLM | 2 | 0.827 | 10 | 1.47 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 3 | 0.826 | 16 | 1.51 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 4 | 0.826 | 7 | 1.42 | Top 8 by both metrics |
| Google Gemma 4 | Adjusted LLM | 5 | 0.826 | 6 | 1.41 | Top 8 by both metrics |
| Google Gemini 3.5 Flash | Adjusted LLM | 6 | 0.825 | 3 | 1.38 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 7 | 0.825 | 9 | 1.46 | Top 8 by COMET-Kiwi |
| Anthropic Claude Fable 5 | Adjusted LLM | 8 | 0.824 | 11 | 1.48 | Top 8 by COMET-Kiwi |
| Translated Lara 2 | Adjusted NMT | 9 | 0.823 | 15 | 1.49 | Real-time place |
| OpenAI GPT-4.1 | Adjusted LLM | 10 | 0.823 | 17 | 1.52 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 11 | 0.823 | 13 | 1.48 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 12 | 0.822 | 1 | 1.33 | Top 8 by MetricX |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 13 | 0.819 | 20 | 1.55 | Not shortlisted |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 14 | 0.818 | 18 | 1.52 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 15 | 0.818 | 12 | 1.48 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 16 | 0.818 | 14 | 1.48 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 17 | 0.816 | 19 | 1.55 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 18 | 0.815 | 21 | 1.55 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 19 | 0.813 | 23 | 1.59 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 20 | 0.813 | 2 | 1.37 | Top 8 by MetricX |
| MiniMax M2.5 | Adjusted LLM | 21 | 0.812 | 22 | 1.57 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 22 | 0.806 | 5 | 1.40 | Top 8 by MetricX |
| ModernMT Enterprise Edition | Adjusted NMT | 23 | 0.803 | 8 | 1.44 | Top 8 by MetricX |
| Microsoft Translator | Off-the-shelf NMT | 24 | 0.800 | 24 | 1.65 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 25 | 0.793 | 26 | 1.79 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 26 | 0.786 | 25 | 1.77 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 27 | 0.782 | 28 | 1.82 | Not shortlisted |
| SYSTRAN PNMT API | Off-the-shelf NMT | 28 | 0.780 | 27 | 1.81 | Not shortlisted |
| Cohere Tiny Aya | Adjusted LLM | 29 | 0.763 | 29 | 2.16 | Not shortlisted |
| Supertext LLM | Off-the-shelf NMT | 30 | 0.753 | 30 | 2.77 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 31 | 0.580 | 31 | 6.10 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.825 | — | 1.39 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.826 | — | 1.36 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.816 | — | 1.50 | 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 31 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 three one-step solutions that meet general translation quality are all on models in the top 8 of COMET-Kiwi. None of the four that made the shortlist through MetricX alone meet it: translations from Cohere Command A Translate, second on MetricX and 20th on COMET-Kiwi, have 80.8 weighted errors per 1,000 words. The models nearest the cut are OpenAI GPT-4.1, 10th on COMET-Kiwi, and xAI Grok 4.20 (reasoning), 12th 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 | $7.50 |
| Standard | OpenAI GPT-5.6 Sol, two-step | $16.90 |
| Advanced | Anthropic Claude Fable 5, three-step with the gender-neutrality agent | $1,034 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.14 | Open-weight LLMs |
| Mistral Large 3 | $0.8 | LLMs |
| Anthropic Claude Haiku 4.5 | $2.13 | LLMs |
| Google Gemini 3.5 Flash | $3.39 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.53 | LLMs |
| Anthropic Claude Sonnet 5 | $5.74 | LLMs |
| Anthropic Claude Sonnet 4.6 | $6.45 | LLMs |
| OpenAI GPT-5.6 Sol | $7.5 | LLMs |
| OpenAI GPT-5.5 | $10.78 | LLMs |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $29.08 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Opus 4.8 | 0.475 | LLMs |
| Anthropic Claude Sonnet 5 | 0.475 | LLMs |
| Anthropic Claude Fable 5 | 0.474 | LLMs |
| Mistral Medium 3.1 | 0.424 | LLMs |
| Mistral Large 3 | 0.416 | LLMs |
| AI21 Jamba Large 1.7 | 0.411 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.375 | Open-weight LLMs |
| Anthropic Claude Haiku 4.5 | 0.37 | LLMs |
| DeepSeek V3.2 | 0.365 | LLMs |
| MoonshotAI Kimi K2.5 | 0.365 | LLMs |
| z.ai GLM-5 | 0.364 | LLMs |
| Cohere Command A Translate | 0.361 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.358 | LLMs |
| MiniMax M2.5 | 0.336 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.335 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.333 | LLMs |
| Cohere Tiny Aya | 0.312 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.312 | LLMs |
| Google Gemma 4 | 0.311 | Open-weight LLMs |
| Google Gemini 3.5 Flash | 0.309 | LLMs |
| OpenAI GPT-5.6 Luna | 0.306 | LLMs |
| OpenAI GPT-5.5 | 0.305 | LLMs |
| OpenAI GPT-5.6 Sol | 0.305 | LLMs |
| OpenAI GPT-4.1 | 0.304 | LLMs |
| OpenAI GPT-5.6 Terra | 0.304 | LLMs |
The cost rises with the steps from one level to the next. The cheapest solution that meets every Basic requirement is the one-step solution on OpenAI GPT-5.6 Sol, at $7.50 per million characters. The cheapest that meets every Standard requirement is the two-step solution (translation + post-edit) on the same model, at $16.90 per million characters. The tokenizers produce a median 0.36 tokens per character of Serbian, 1.4 times as many as English text, with 0.305 for OpenAI GPT-5.6 Sol and 0.474 for Anthropic Claude Fable 5.
| 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 | 60 |
| Standard | Anthropic Claude Fable 5, one-step LLM adjusted for Standard requirements | 62 |
| Advanced | Anthropic Claude Fable 5, three-step with the gender-neutrality agent | 224 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.2 | MT engines |
| ModernMT Enterprise Edition | 1 | MT engines |
| Translated Lara 2 | 1.3 | LLM-based MT engines |
| Supertext LLM | 1.6 | LLM-based MT engines |
| SYSTRAN PNMT API | 2.1 | MT engines |
| OpenAI gpt-oss-120b | 2.3 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 6.1 | MT engines |
| Google Gemini 3.5 Flash | 14 | LLMs |
| OpenAI GPT-5.6 Luna | 14 | LLMs |
| Cohere Tiny Aya | 15 | LLMs |
| OpenAI GPT-4.1 | 21 | LLMs |
| OpenAI GPT-5.6 Terra | 26 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 26 | LLMs |
| Anthropic Claude Haiku 4.5 | 33 | LLMs |
| DeepSeek V3.2 | 40 | LLMs |
| MiniMax M2.5 | 40 | LLMs |
| Mistral Medium 3.1 | 46 | LLMs |
| Cohere Command A Translate | 50 | LLMs |
| Swiss AI Apertus-8B-2509 | 53 | Open-weight LLMs |
| Anthropic Claude Sonnet 5 | 56 | LLMs |
| Mistral Large 3 | 56 | LLMs |
| AI21 Jamba Large 1.7 | 56 | LLMs |
| OpenAI GPT-5.6 Sol | 60 | LLMs |
| Anthropic Claude Fable 5 | 62 | LLMs |
| Anthropic Claude Sonnet 4.6 | 64 | LLMs |
| OpenAI GPT-5.5 | 68 | LLMs |
| Anthropic Claude Opus 4.8 | 73 | LLMs |
| xAI Grok 4.20 (reasoning) | 83 | LLMs |
| AppTek NMT | 93 | MT engines |
| Google Gemini 3.1 Pro Preview | 113 | LLMs |
| Alibaba Qwen3-235B-A22B | 154 | Open-weight LLMs |
| Google Gemma 4 | 202 | Open-weight LLMs |
| z.ai GLM-5 | 251 | LLMs |
| MoonshotAI Kimi K2.5 | 276 | 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 Serbian.
Translations are real-time (<2 sec) from five models, four MT engines or LLM-based MT engines and the open-weight OpenAI gpt-oss-120b, which did not make the shortlist. The fastest solution that meets every Basic requirement takes 60 seconds per 1,000 words, the fastest that meets every Standard requirement 62, and the only one that meets every Advanced requirement 224.
| 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% | ✓ | $1,034 | 224 |
| OpenAI GPT-5.6 Sol | two-step, with the gender-neutrality agent | 98.2% | ✗ | $18.58 | 103 |
For tone of voice in Serbian we check the formal register, with Vi as the form of address, and for gender neutrality we check neutral rephrasing with impersonal constructions. None of the shortlisted one-step solutions meet gender neutrality. They work without the gender-neutrality agent and have a median of 7.2 gender-neutrality errors per 1,000 words. In the translations from the three-step solution (translation + LQA + repair) on Anthropic Claude Fable 5 with the agent, 100% of the sentences about people are gender-neutral.
The similarity pair of automatic metrics falls short by a narrow margin, with an agreement of 0.460 against a threshold of 0.461. The quality-estimation pair clears its threshold, 0.57 against 0.53. The three LQA judges agree on a median of 96.0% of the sentences of the shortlisted solutions, so the Intento LQA results on this page rest on consistent assessments.
Leave feedback