The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Amharic 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 Amharic. |
| Tone of voice | The formal register (respectful verb 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 Amharic, 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 11 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 solutions with one or two steps meet the Basic requirements in Amharic, which cover general translation quality alone. For the Standard requirements, only the three-step solution (translation + LQA + repair) on Google Gemini 3.1 Pro Preview fits a cost cap from $613 to $650 per million characters or a time limit from 317 to under 485 seconds per 1,000 words. None fit a cap of at most $612 or a limit of at most 316 seconds. None of the solutions that meet every Advanced requirement cost under $1,043 per million characters.
For the Advanced requirements, the gender-neutrality agent works with both three-step solutions (translation + LQA + repair). On Google Gemini 3.1 Pro Preview the cost goes from $612 to $1,043 per million characters and the time from 316 to 321 seconds per 1,000 words. On OpenAI GPT-5.5 the cost goes from $651 to $1,079 per million characters and the time from 485 to 484 seconds per 1,000 words. Total errors fall from 11.6 to 11.5 and from 18.8 to 15.8 per 1,000 words.
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 with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
three-step with the gender-neutrality agent
Google Gemini 3.1 Pro Preview
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 2: two or more solutions, none simple | ☆☆ | three-step | Google Gemini 3.1 Pro Preview | 11.6 | $612 | 316 |
| Standard | Tier 2: two or more solutions, none simple | ☆☆ | three-step | Google Gemini 3.1 Pro Preview | 11.6 | $612 | 316 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | with the gender-neutrality agent | Google Gemini 3.1 Pro Preview | 11.5 | $1,043 | 321 |
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.
Neither of the two three-step solutions (translation + LQA + repair) that meet every Basic requirement in Amharic, on Google Gemini 3.1 Pro Preview and OpenAI GPT-5.5, drops out at the Standard requirements: translations from both have 100% of the sentences in the tone of voice and 100% of the tags in place, and both meet formatting.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.16 | 104.5 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.97 | 54.4 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 5.42 | 38.6 | Misses a Standard requirement |
| Anthropic Claude Sonnet 5 | 8.36 | 68.2 | Misses a Standard requirement |
| Microsoft Translator | 10 | 147 | Misses a Standard requirement |
| Google Translation LLM | 18.66 | 105.8 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 20.88 | 23.1 | Misses a Standard requirement |
| Anthropic Claude Opus 4.8 | 21.39 | 55.4 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 81.1 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 30.41 | 34.3 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 42.49 | 27 | Misses a Standard requirement |
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 55.4 |
| Gender neutrality | 1.7 |
| Tag handling | 0.7 |
| Tone of voice | 0.3 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
General translation quality is also what the two-step solutions (translation + post-edit) miss in Amharic: those on OpenAI GPT-5.5 and Google Gemini 3.1 Pro Preview meet tone of voice, tag handling and formatting, but have 35.6 and 37.6 weighted errors per 1,000 words, over the bar.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Amharic | Class |
|---|---|---|
| Microsoft Translator | 147 | MT engines |
| Google Translation LLM | 105.8 | LLM-based MT engines |
| Google Gemma 4 | 104.5 | Open-weight LLMs |
| Translated Lara 2 | 81.1 | LLM-based MT engines |
| Anthropic Claude Sonnet 5 | 68.2 | LLMs |
| Anthropic Claude Opus 4.8 | 55.4 | LLMs |
| Google Gemini 3.5 Flash | 54.4 | LLMs |
| Google Gemini 3.1 Pro Preview | 38.6 | LLMs |
| OpenAI GPT-5.5 | 34.3 | LLMs |
| Anthropic Claude Fable 5 | 27 | LLMs |
| OpenAI GPT-5.6 Sol | 23.1 | LLMs |
No translation into Amharic: AI21 Jamba Large 1.7 (garbage or near-empty), ALL-Lab (far too short, wrong language, wrong script), Swiss AI Apertus-8B-2509 (garbage or near-empty), Cohere Command A Translate (garbage or near-empty).
Translations from the seven shortlisted LLMs other than the open-weight Google Gemma 4 have 23.1 to 68.2 weighted errors per 1,000 words in one step in Amharic. Translations from the three shortlisted MT engines, two of them LLM-based, have 81.1 to 147.0 weighted errors per 1,000 words. Those from Google Translation LLM have 11 critical errors, the most on the shortlist. The three shortlisted LLMs that do not meet our language-support criteria are Google Gemini 3.5 Flash, Anthropic Claude Sonnet 5 and Google Gemma 4.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| DeepSeek V3.2 | 0.743 | 10 | 1.49 | 9 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.734 | 12 | 1.54 | 10 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.726 | 13 | 1.75 | 14 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.717 | 15 | 1.71 | 13 | No | Not shortlisted |
| z.ai GLM-5 | 0.715 | 16 | 1.91 | 16 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.709 | 17 | 2.38 | 19 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.695 | 18 | 2.24 | 17 | No | Not shortlisted |
| OpenAI GPT-4.1 | 0.681 | 19 | 2.28 | 18 | No | Not shortlisted |
| MiniMax M2.5 | 0.672 | 20 | 2.83 | 22 | No | Not shortlisted |
| Mistral Large 3 | 0.657 | 21 | 2.66 | 21 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.655 | 22 | 2.63 | 20 | No | Not shortlisted |
| Amazon Translate | 0.649 | 23 | 3.04 | 23 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.647 | 24 | 3.44 | 24 | No | Not shortlisted |
| Cohere Tiny Aya | 0.603 | 25 | 4.51 | 25 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.580 | 26 | 5.06 | 26 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.501 | 27 | 6.89 | 27 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.461 | 28 | 8.23 | 28 | No | Not shortlisted |
| Google Gemini 3.1 Pro Preview | 0.758 | 1 | 1.24 | 2 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.758 | 2 | 1.14 | 1 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | 0.757 | 3 | 1.27 | 3 | Yes | Top 8 by both metrics |
| Google Gemini 3.5 Flash | 0.751 | 4 | 1.46 | 8 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.750 | 5 | 1.33 | 5 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.5 | 0.749 | 6 | 1.39 | 7 | Yes | Top 8 by both metrics |
| Google Translation LLM | 0.749 | 7 | 1.86 | 15 | Yes | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | 0.748 | 8 | 1.31 | 4 | Yes | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | 0.748 | 9 | 1.36 | 6 | Yes | Top 8 by MetricX |
| Translated Lara 2 | 0.740 | 11 | 1.54 | 11 | Yes | Real-time place |
| Microsoft Translator | 0.719 | 14 | 1.70 | 12 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 1 | 0.758 | 2 | 1.24 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 2 | 0.758 | 1 | 1.14 | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | Adjusted LLM | 3 | 0.757 | 3 | 1.27 | Top 8 by both metrics |
| Google Gemini 3.5 Flash | Adjusted LLM | 4 | 0.751 | 8 | 1.46 | Top 8 by both metrics |
| Google Gemma 4 | Adjusted LLM | 5 | 0.750 | 5 | 1.33 | Top 8 by both metrics |
| OpenAI GPT-5.5 | Adjusted LLM | 6 | 0.749 | 7 | 1.39 | Top 8 by both metrics |
| Google Translation LLM | Adjusted NMT | 7 | 0.749 | 15 | 1.86 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 8 | 0.748 | 4 | 1.31 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 9 | 0.748 | 6 | 1.36 | Top 8 by MetricX |
| DeepSeek V3.2 | Adjusted LLM | 10 | 0.743 | 9 | 1.49 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 11 | 0.740 | 11 | 1.54 | Real-time place |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 12 | 0.734 | 10 | 1.54 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 13 | 0.726 | 14 | 1.75 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 14 | 0.719 | 12 | 1.70 | Real-time place |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 15 | 0.717 | 13 | 1.71 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 16 | 0.715 | 16 | 1.91 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 17 | 0.709 | 19 | 2.38 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 18 | 0.695 | 17 | 2.24 | Not shortlisted |
| OpenAI GPT-4.1 | Adjusted LLM | 19 | 0.681 | 18 | 2.28 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 20 | 0.672 | 22 | 2.83 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 21 | 0.657 | 21 | 2.66 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Off-the-shelf LLM | 22 | 0.655 | 20 | 2.63 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 23 | 0.649 | 23 | 3.04 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 24 | 0.647 | 24 | 3.44 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 25 | 0.603 | 25 | 4.51 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 26 | 0.580 | 26 | 5.06 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 27 | 0.501 | 27 | 6.89 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 28 | 0.461 | 28 | 8.23 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.748 | — | 1.35 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.754 | — | 1.29 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.719 | — | 1.69 | 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 28 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 one-step solution on Google Gemini 3.1 Pro Preview, first on COMET-Kiwi and second on MetricX, has 38.6 weighted errors per 1,000 words by Intento LQA, and the one on OpenAI GPT-5.6 Sol, third on both metrics, has 23.1. Google Translation LLM is on the shortlist through COMET-Kiwi alone, seventh there and 15th on MetricX. The one-step solution on that model has 105.8 weighted errors per 1,000 words. Nearest the cut in the tail is DeepSeek V3.2, tenth on COMET-Kiwi and ninth on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | Google Gemini 3.1 Pro Preview, three-step | $612 |
| Standard | Google Gemini 3.1 Pro Preview, three-step | $612 |
| Advanced | Google Gemini 3.1 Pro Preview, three-step with the gender-neutrality agent | $1,043 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.16 | Open-weight LLMs |
| Google Gemini 3.5 Flash | $3.97 | LLMs |
| Google Gemini 3.1 Pro Preview | $5.42 | LLMs |
| Anthropic Claude Sonnet 5 | $8.36 | LLMs |
| Microsoft Translator | $10 | MT engines |
| Google Translation LLM | $18.66 | LLM-based MT engines |
| OpenAI GPT-5.6 Sol | $20.88 | LLMs |
| Anthropic Claude Opus 4.8 | $21.39 | LLMs |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| OpenAI GPT-5.5 | $30.41 | LLMs |
| Anthropic Claude Fable 5 | $42.49 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Mistral Medium 3.1 | 1.962 | LLMs |
| z.ai GLM-5 | 1.707 | LLMs |
| Mistral Large 3 | 1.608 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 1.566 | LLMs |
| xAI Grok 4.20 (reasoning) | 1.516 | LLMs |
| DeepSeek V3.2 | 1.38 | LLMs |
| OpenAI GPT-5.6 Sol | 1.287 | LLMs |
| OpenAI GPT-4.1 | 1.262 | LLMs |
| OpenAI GPT-5.6 Luna | 1.259 | LLMs |
| OpenAI GPT-5.6 Terra | 1.246 | LLMs |
| OpenAI GPT-5.5 | 1.238 | LLMs |
| MiniMax M2.5 | 1.137 | LLMs |
| Anthropic Claude Fable 5 | 0.991 | LLMs |
| Anthropic Claude Opus 4.8 | 0.983 | LLMs |
| Anthropic Claude Sonnet 5 | 0.977 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.965 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.963 | LLMs |
| MoonshotAI Kimi K2.5 | 0.96 | LLMs |
| Cohere Tiny Aya | 0.56 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.504 | LLMs |
| Google Gemma 4 | 0.494 | Open-weight LLMs |
| Google Gemini 3.5 Flash | 0.491 | LLMs |
| 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 | 316 |
| Standard | Google Gemini 3.1 Pro Preview, three-step | 316 |
| Advanced | Google Gemini 3.1 Pro Preview, three-step with the gender-neutrality agent | 321 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Google Cloud Advanced Translation API | 0.2 | MT engines |
| Microsoft Translator | 0.3 | MT engines |
| ModernMT Enterprise Edition | 0.5 | MT engines |
| Google Translation LLM | 1.4 | LLM-based MT engines |
| Amazon Translate | 1.5 | MT engines |
| Yandex Cloud Translate v2 | 2.8 | MT engines |
| Translated Lara 2 | 5.2 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 6.7 | Open-weight LLMs |
| Alibaba Machine Translation API | 26 | MT engines |
| Cohere Tiny Aya | 38 | LLMs |
| OpenAI GPT-5.6 Luna | 49 | LLMs |
| DeepSeek V3.2 | 51 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 61 | LLMs |
| Google Gemini 3.5 Flash | 68 | LLMs |
| Anthropic Claude Haiku 4.5 | 73 | LLMs |
| OpenAI GPT-4.1 | 75 | LLMs |
| Google Gemini 3.1 Pro Preview | 93 | LLMs |
| OpenAI GPT-5.6 Terra | 104 | LLMs |
| MiniMax M2.5 | 124 | LLMs |
| xAI Grok 4.20 (reasoning) | 138 | LLMs |
| Anthropic Claude Sonnet 4.6 | 154 | LLMs |
| Anthropic Claude Sonnet 5 | 172 | LLMs |
| Google Gemma 4 | 180 | Open-weight LLMs |
| Anthropic Claude Fable 5 | 200 | LLMs |
| OpenAI GPT-5.6 Sol | 201 | LLMs |
| OpenAI GPT-5.5 | 216 | LLMs |
| Mistral Large 3 | 220 | LLMs |
| Anthropic Claude Opus 4.8 | 226 | LLMs |
| MoonshotAI Kimi K2.5 | 262 | LLMs |
| Mistral Medium 3.1 | 337 | LLMs |
| z.ai GLM-5 | 508 | 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 Amharic.
In Amharic, the one-step solutions take 68 to 226 seconds per 1,000 words on the eight shortlisted LLMs, against 0.3 to 5.2 on the three shortlisted MT engines.
| 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, with the gender-neutrality agent | 100% | ✓ | $1,043 | 321 |
| OpenAI GPT-5.5 | three-step, with the gender-neutrality agent | 100% | ✓ | $1,079 | 484 |
For Amharic we check the formal register, with respectful verb forms. For gender neutrality we check neutral rephrasing through plural, nominal or infinitive forms, and a default gender is not allowed. The other options in Amharic are a masculine default and double (slash) forms. Both three-step solutions (translation + LQA + repair) meet tone of voice, and with the gender-neutrality agent 100% of the sentences about people in their translations are gender-neutral.
The quality-estimation pair of automatic metrics has an agreement of 0.68 in Amharic, against a threshold of 0.53, and the similarity pair 0.57, against a threshold of 0.46. For the 11 shortlisted solutions, the median share of sentences on which the three LQA judges agree is 92.1%. With agreement at these levels, the shortlist and the Intento LQA error counts on this page carry weight.
Leave feedback