The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Cebuano 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 Cebuano. |
| Tone of voice | The formal register (respectful forms) 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 Cebuano, 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. Inconclusive: the three LQA judges agree on fewer than 90% of the sentences of the solution, so it has no general translation quality result (09-02). Gender neutrality: the Advanced requirements count the two best solutions of the language, with the gender-neutrality agent where they need it. The results of the other solutions are shown for information.
Two three-step solutions (translation + LQA + repair) meet every Standard requirement in Cebuano, and the dearer one has more errors. The one on OpenAI GPT-5.5 costs $631 per million characters, against $604 for the one on Google Gemini 3.1 Pro Preview. It has 11.5 total errors per 1,000 words against 11.2, and takes 306 seconds per 1,000 words against 293. Optimizing for quality, cost or latency selects the one on Google Gemini 3.1 Pro Preview, under the Basic, Standard and Advanced requirements alike.
The three-step solution (translation + LQA + repair) on Anthropic Claude Fable 5 meets every Standard requirement in Cebuano except general translation quality. It has 20.4 weighted errors per 1,000 words, just over the bar. It takes 247 seconds per 1,000 words, less than either solution that meets the requirements, and costs $672 per million characters, more than either.
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.1 Pro Preview
three-step
Google Gemini 3.1 Pro Preview
three-step
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 | ☆☆ | three-step | Google Gemini 3.1 Pro Preview | 11.2 | $604 | 293 |
| Standard | Tier 2: two or more solutions, none simple | ☆☆ | three-step | Google Gemini 3.1 Pro Preview | 11.2 | $604 | 293 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | three-step | Google Gemini 3.1 Pro Preview | 11.2 | $604 | 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.
Two three-step solutions (translation + LQA + repair) meet the Basic requirements in Cebuano, those on Google Gemini 3.1 Pro Preview and OpenAI GPT-5.5. Both also meet the tone of voice, tag handling and formatting of the Standard requirements, and the gender neutrality of the Advanced ones without the gender-neutrality agent. None of the one-step or two-step solutions meet even the Basic requirements. Those with a conclusive Intento LQA result have 34.3 or more weighted errors per 1,000 words, so each level takes three steps.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.15 | 61.3 | Misses a Standard requirement |
| xAI Grok 4.20 (reasoning) | 1.23 | 118 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 5.06 | 34.9 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 11.62 | 38.6 | Misses a Standard requirement |
| Google Cloud Advanced Translation API | 20 | 95.9 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 32.92 | 39.2 | Misses a Standard requirement |
Among the one-step solutions in Cebuano with a conclusive Intento LQA result, paying more than $5.06 per million characters does not buy fewer weighted errors. At that price the one on Google Gemini 3.1 Pro Preview has the fewest weighted errors, 34.9 per 1,000 words. The three dearer ones, on OpenAI GPT-5.5, Google Cloud Advanced Translation API and Anthropic Claude Fable 5, cost $11.62 to $32.92 per million characters and have 38.6 to 95.9 weighted errors per 1,000 words.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 50.2 |
| Tone of voice | 7.8 |
| Tag handling | 0 |
| Gender neutrality | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
General translation quality is the requirement that takes the third step in Cebuano. Two steps are enough for tone of voice on two models: the two-step solutions (translation + post-edit) on OpenAI GPT-5.5 and Google Gemma 4 have 100% of the sentences in the formal register, against 95.7% and 19.1% in one step. Only the three-step solutions (translation + LQA + repair) on Google Gemini 3.1 Pro Preview and OpenAI GPT-5.5 meet general translation quality, with 11.2 and 11.5 weighted errors per 1,000 words.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Cebuano | Class |
|---|---|---|
| xAI Grok 4.20 (reasoning) | 118 | LLMs |
| Google Cloud Advanced Translation API | 95.9 | MT engines |
| Google Gemma 4 | 61.3 | Open-weight LLMs |
| Anthropic Claude Fable 5 | 39.2 | LLMs |
| OpenAI GPT-5.5 | 38.6 | LLMs |
| Google Gemini 3.1 Pro Preview | 34.9 | LLMs |
In one step, translations from the three shortlisted LLMs that meet our language-support criteria, Google Gemini 3.1 Pro Preview, OpenAI GPT-5.5 and Anthropic Claude Fable 5, have 34.9 to 39.2 weighted errors per 1,000 words. Those from the open-weight Google Gemma 4 and from xAI Grok 4.20 (reasoning) have 61.3 and 118.0 weighted errors per 1,000 words, over the 60 the criteria allow. Of the one-step solutions with a conclusive result, only the one on the MT engine Google Cloud Advanced Translation API has critical errors, 7 of them.
| Rank | Model | MetricX (lower is better) | 83% confidence interval, from | to | Status |
|---|---|---|---|---|---|
| 1 | OpenAI GPT-5.6 Terra | 1.96 | 1.75 | 2.17 | Shortlisted |
| 2 | Google Gemini 3.1 Pro Preview | 2.11 | 1.89 | 2.32 | Shortlisted |
| 3 | OpenAI GPT-5.5 | 2.21 | 1.98 | 2.43 | Shortlisted |
| 4 | Google Gemma 4 | 2.21 | 1.99 | 2.44 | Shortlisted |
| 5 | Google Gemini 3.5 Flash | 2.28 | 2.05 | 2.51 | Shortlisted |
| 6 | Anthropic Claude Opus 4.8 | 2.31 | 2.10 | 2.53 | Shortlisted |
| 7 | Anthropic Claude Sonnet 4.6 | 2.43 | 2.20 | 2.65 | Shortlisted |
| 8 | Anthropic Claude Fable 5 | 2.49 | 2.23 | 2.75 | Shortlisted |
| 9 | Translated Lara 2 | 2.51 | 2.28 | 2.74 | Shortlisted |
| 10 | xAI Grok 4.20 (reasoning) | 2.61 | 2.37 | 2.86 | Shortlisted |
| 11 | Mistral Large 3 | 2.63 | 2.40 | 2.86 | Not shortlisted |
| 12 | z.ai GLM-5 | 2.63 | 2.39 | 2.88 | Not shortlisted |
| 13 | OpenAI GPT-4.1 | 2.65 | 2.40 | 2.90 | Not shortlisted |
| 14 | MoonshotAI Kimi K2.5 | 2.77 | 2.51 | 3.03 | Not shortlisted |
| 15 | DeepSeek V3.2 | 2.77 | 2.53 | 3.02 | Not shortlisted |
| 16 | Anthropic Claude Sonnet 5 | 2.77 | 2.54 | 3.01 | Not shortlisted |
| 17 | OpenAI gpt-oss-120b | 3.05 | 2.75 | 3.35 | Shortlisted |
| 18 | Google Cloud Advanced Translation API | 3.14 | 2.63 | 3.64 | Shortlisted |
| 19 | Mistral Medium 3.1 | 3.64 | 3.35 | 3.93 | Not shortlisted |
| 20 | Cohere Command A Translate | 3.64 | 3.37 | 3.91 | Not shortlisted |
| 21 | Anthropic Claude Haiku 4.5 | 3.69 | 3.37 | 4.01 | Not shortlisted |
| 22 | Alibaba Qwen3-235B-A22B | 3.78 | 3.45 | 4.12 | Not shortlisted |
| 23 | MiniMax M2.5 | 4.36 | 4.04 | 4.67 | Not shortlisted |
| 24 | Yandex Cloud Translate v2 | 5.01 | 4.59 | 5.42 | Not shortlisted |
| 25 | AI21 Jamba Large 1.7 | 5.29 | 4.88 | 5.69 | Not shortlisted |
| 26 | Swiss AI Apertus-8B-2509 | 5.82 | 5.23 | 6.41 | Not shortlisted |
| Model | Solution complexity | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|
| OpenAI GPT-5.6 Terra | Adjusted LLM | 1 | 1.96 | Top 10 by MetricX |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 2 | 2.11 | Top 10 by MetricX |
| OpenAI GPT-5.5 | Adjusted LLM | 3 | 2.21 | Top 10 by MetricX |
| Google Gemma 4 | Adjusted LLM | 4 | 2.21 | Top 10 by MetricX |
| Google Gemini 3.5 Flash | Adjusted LLM | 5 | 2.28 | Top 10 by MetricX |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 6 | 2.31 | Top 10 by MetricX |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 7 | 2.43 | Top 10 by MetricX |
| Anthropic Claude Fable 5 | Adjusted LLM | 8 | 2.49 | Top 10 by MetricX |
| Translated Lara 2 | Adjusted NMT | 9 | 2.51 | Top 10 by MetricX |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 10 | 2.61 | Top 10 by MetricX |
| Mistral Large 3 | Adjusted LLM | 11 | 2.63 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 12 | 2.63 | Not shortlisted |
| OpenAI GPT-4.1 | Adjusted LLM | 13 | 2.65 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 14 | 2.77 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 15 | 2.77 | Not shortlisted |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 16 | 2.77 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 17 | 3.05 | Real-time place |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 18 | 3.14 | Real-time place |
| Mistral Medium 3.1 | Adjusted LLM | 19 | 3.64 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 20 | 3.64 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 21 | 3.69 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 22 | 3.78 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 23 | 4.36 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 24 | 5.01 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 25 | 5.29 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 26 | 5.82 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 2.13 | Same model family as OpenAI GPT-5.6 Terra |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 1.98 | Same model family as OpenAI GPT-5.6 Terra |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 2.73 | 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. Ranks among 26 models, with each model family counted once, and the top 10 in bold. Lower is better on MetricX. The scores rank models within this language and are not a requirement result.
MetricX ranks Google Gemma 4 fourth in Cebuano and Anthropic Claude Fable 5 eighth, but by Intento LQA the one-step solution on the latter has fewer weighted errors, 39.2 per 1,000 words against 61.3. The one on xAI Grok 4.20 (reasoning), last of the ten MetricX places, has the most weighted errors of the six with a conclusive result, 118.0 per 1,000 words. Nearest the cut are Mistral Large 3 and z.ai GLM-5, 11th and 12th. OpenAI gpt-oss-120b and Google Cloud Advanced Translation API, 17th and 18th, hold the real-time places.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | Google Gemini 3.1 Pro Preview, three-step | $604 |
| Standard | Google Gemini 3.1 Pro Preview, three-step | $604 |
| Advanced | Google Gemini 3.1 Pro Preview, three-step | $604 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.15 | Open-weight LLMs |
| xAI Grok 4.20 (reasoning) | $1.23 | LLMs |
| Google Gemini 3.5 Flash | $3.74 | LLMs |
| OpenAI GPT-5.6 Terra | $4.64 | LLMs |
| Google Gemini 3.1 Pro Preview | $5.06 | LLMs |
| Anthropic Claude Sonnet 4.6 | $7.28 | LLMs |
| OpenAI GPT-5.5 | $11.62 | LLMs |
| Anthropic Claude Opus 4.8 | $16.86 | LLMs |
| Google Cloud Advanced Translation API | $20 | MT engines |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $32.92 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Opus 4.8 | 0.496 | LLMs |
| Anthropic Claude Sonnet 5 | 0.493 | LLMs |
| Anthropic Claude Fable 5 | 0.49 | LLMs |
| Mistral Medium 3.1 | 0.402 | LLMs |
| Mistral Large 3 | 0.392 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.357 | LLMs |
| AI21 Jamba Large 1.7 | 0.354 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.345 | LLMs |
| Cohere Command A Translate | 0.343 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.329 | Open-weight LLMs |
| z.ai GLM-5 | 0.328 | LLMs |
| DeepSeek V3.2 | 0.32 | LLMs |
| MoonshotAI Kimi K2.5 | 0.314 | LLMs |
| Google Gemini 3.5 Flash | 0.312 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.31 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.308 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.308 | LLMs |
| Google Gemma 4 | 0.305 | Open-weight LLMs |
| MiniMax M2.5 | 0.297 | LLMs |
| OpenAI GPT-5.6 Sol | 0.294 | LLMs |
| OpenAI GPT-5.6 Luna | 0.292 | LLMs |
| OpenAI GPT-5.6 Terra | 0.292 | LLMs |
| OpenAI GPT-5.5 | 0.289 | LLMs |
| OpenAI GPT-4.1 | 0.285 | LLMs |
On one model the cost follows the steps: the solutions on Google Gemini 3.1 Pro Preview cost $5.06 per million characters in one step, $32.55 in two steps (translation + post-edit) and $604 in three (translation + LQA + repair). Tokens per character do not set the cost alone. The tokenizer of Anthropic Claude Opus 4.8 produces 0.496 tokens per character and that of Anthropic Claude Fable 5 0.49, yet the one-step solutions on them cost $16.86 and $32.92 per million characters. From one level to the next the cost stays the same.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | Google Gemini 3.1 Pro Preview, three-step | 293 |
| Standard | Google Gemini 3.1 Pro Preview, three-step | 293 |
| Advanced | Google Gemini 3.1 Pro Preview, three-step | 293 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Google Cloud Advanced Translation API | 0.2 | MT engines |
| Translated Lara 2 | 1.5 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 2.6 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 3 | MT engines |
| OpenAI GPT-4.1 | 19 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 23 | LLMs |
| OpenAI GPT-5.6 Luna | 25 | LLMs |
| Anthropic Claude Haiku 4.5 | 36 | LLMs |
| OpenAI GPT-5.6 Terra | 39 | LLMs |
| Mistral Medium 3.1 | 45 | LLMs |
| MiniMax M2.5 | 52 | LLMs |
| AI21 Jamba Large 1.7 | 56 | LLMs |
| Cohere Command A Translate | 59 | LLMs |
| Swiss AI Apertus-8B-2509 | 60 | Open-weight LLMs |
| xAI Grok 4.20 (reasoning) | 66 | LLMs |
| Google Gemini 3.5 Flash | 68 | LLMs |
| OpenAI GPT-5.6 Sol | 70 | LLMs |
| Mistral Large 3 | 75 | LLMs |
| DeepSeek V3.2 | 76 | LLMs |
| Anthropic Claude Sonnet 5 | 81 | LLMs |
| Google Gemini 3.1 Pro Preview | 82 | LLMs |
| Anthropic Claude Fable 5 | 84 | LLMs |
| Anthropic Claude Sonnet 4.6 | 87 | LLMs |
| OpenAI GPT-5.5 | 96 | LLMs |
| Anthropic Claude Opus 4.8 | 104 | LLMs |
| Alibaba Qwen3-235B-A22B | 133 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 199 | LLMs |
| Google Gemma 4 | 208 | Open-weight LLMs |
| z.ai GLM-5 | 225 | 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 Cebuano.
The real-time models in Cebuano are an MT engine, an LLM-based MT engine and an open-weight LLM: Google Cloud Advanced Translation API, Translated Lara 2 and OpenAI gpt-oss-120b. Translations from them have 6.38% to 42.6% of the sentences in the formal register. Only the solution on Google Cloud Advanced Translation API has a conclusive Intento LQA result: 95.9 weighted errors per 1,000 words. The three-step solutions (translation + LQA + repair) that meet the requirements take 293 and 306 seconds per 1,000 words at each level.
| 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% | ✓ | $604 | 293 |
| OpenAI GPT-5.5 | three-step | 100% | ✓ | $631 | 306 |
For Cebuano we check the formal register, with respectful forms of address, and gender-neutral wording with no default gender. Cebuano has no grammatical gender for persons, so translations are neutral by default and the requirement comes down to avoiding gendered terms. In one step, the shortlisted solutions have 4.3% to 97.9% of the sentences in the formal register, against a bar of 100%. Without the gender-neutrality agent, translations from the two best solutions have 100% of the sentences about people gender-neutral.
For the 12 shortlisted solutions in Cebuano, the median share of sentences on which the three LQA judges agree is 89.8%. There is no conclusive Intento LQA result for the one-step solution on OpenAI GPT-5.6 Terra, first on MetricX, or for five other shortlisted ones. Every two-step and three-step solution on this page runs on a model with a conclusive one-step result. The similarity pair of automatic metrics agrees at 0.52, with an interval of 0.51 to 0.54 that stays above its threshold of 0.46.
Leave feedback