The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Khmer 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 Khmer. |
| Tone of voice | The formal register (honorific vocabulary) 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 Khmer, 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.
Under the Basic requirements, the two one-step solutions trade cost against errors. The one on OpenAI GPT-5.6 Sol costs $12.04 per million characters and has 19.8 total errors per 1,000 words, the one on OpenAI GPT-5.5 $17.75 and 19.7. For the Standard requirements, the two-step solution (translation + post-edit) on OpenAI GPT-5.6 Sol takes its place, with 17.1 total errors per 1,000 words, $27.09 per million characters and 597 seconds per 1,000 words. Optimize for then selects it for quality, and OpenAI GPT-5.5 for cost and latency.
A cost cap leaves a solution that meets every Standard requirement from $17.75 per million characters up, and one that meets every Advanced requirement from $27.43 up. A time limit leaves them from 216 and 433 seconds per 1,000 words up. The 216 seconds are the time of the one-step solution on OpenAI GPT-5.5, which a limit to one-step solutions leaves alone for the Standard requirements. None of the three solutions added to the shortlist for real time meet general translation quality, tone of voice, tag handling or formatting.
Both of the two best solutions meet the Advanced requirements only with the gender-neutrality agent. The cheaper one, the two-step solution (translation + post-edit) on OpenAI GPT-5.6 Sol with the agent, costs $27.43 per million characters and takes 432 seconds per 1,000 words, against $27.09 and 597 seconds without it. The other, the three-step solution (translation + LQA + repair) on OpenAI GPT-5.5 with the agent, costs $1,021 per million characters and has 7.3 total errors per 1,000 words, against 18.5 for the cheaper one.
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
two-step
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
three-step with the gender-neutrality agent
OpenAI GPT-5.6 Sol
two-step with the gender-neutrality agent
OpenAI GPT-5.6 Sol
two-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 | 19.8 | $12.04 | 253 |
| Standard | Tier 1: simple solution + backup | ★ ☆ | one-step | OpenAI GPT-5.5 | 19.7 | $17.75 | 216 |
| Advanced | Tier 2: two or more solutions, none simple | ☆☆ | with the gender-neutrality agent | OpenAI GPT-5.6 Sol | 18.5 | $27.43 | 432 |
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.
Each level up removes one of the one-step solutions in Khmer. Tone of voice removes the one on OpenAI GPT-5.6 Sol, and the two-step solution (translation + post-edit) on that model replaces it. Gender neutrality removes the one on OpenAI GPT-5.5, so the solutions on both models meet every Advanced requirement only with the gender-neutrality agent. Optimize for selects OpenAI GPT-5.6 Sol for cost under the Basic requirements, for quality under the Standard ones and for cost and latency under the Advanced ones, otherwise OpenAI GPT-5.5.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.17 | 53.4 | Misses a Standard requirement |
| z.ai GLM-5 | 4.19 | 71.5 | Misses a Standard requirement |
| Microsoft Translator | 10 | 127.9 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 12.04 | 18.5 | Misses a Standard requirement |
| Anthropic Claude Sonnet 4.6 | 13.81 | 60 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 17.75 | 19.4 | Meets every Standard requirement |
| Google Translation LLM | 21.15 | 83.4 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 85.7 | Misses a Standard requirement |
| Anthropic Claude Opus 4.8 | 25.65 | 34.3 | Misses a Standard requirement |
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 60 |
| Tone of voice | 0.6 |
| Tag handling | 0.3 |
| Gender neutrality | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
General translation quality is the Standard requirement the fewest one-step solutions meet in Khmer. Of the one-step solutions that do not meet it, the one on Anthropic Claude Opus 4.8 has the fewest weighted errors, 34.3 per 1,000 words. Two-step solutions (translation + post-edit) bring the one on OpenAI GPT-5.6 Sol to 100% of the sentences in the formal register, and those on Anthropic Claude Opus 4.8 and z.ai GLM-5 to 33.9 and 68.2 weighted errors per 1,000 words.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Khmer | Class |
|---|---|---|
| Microsoft Translator | 127.9 | MT engines |
| Translated Lara 2 | 85.7 | LLM-based MT engines |
| Google Translation LLM | 83.4 | LLM-based MT engines |
| z.ai GLM-5 | 71.5 | LLMs |
| Anthropic Claude Sonnet 4.6 | 60 | LLMs |
| Google Gemma 4 | 53.4 | Open-weight LLMs |
| Anthropic Claude Opus 4.8 | 34.3 | LLMs |
| OpenAI GPT-5.5 | 19.4 | LLMs |
| OpenAI GPT-5.6 Sol | 18.5 | LLMs |
No translation into Khmer: AI21 Jamba Large 1.7 (garbage or near-empty, wrong language, wrong script, cut off part-way), Swiss AI Apertus-8B-2509 (cut off part-way).
The one-step solutions on the five LLMs with a conclusive Intento LQA result have 18.5 to 71.5 weighted errors per 1,000 words. The one on the open-weight LLM Google Gemma 4 has 53.4, those on the LLM-based MT engines 83.4 and 85.7, and the one on the MT engine Microsoft Translator 127.9 weighted errors per 1,000 words. Of these six LLMs, OpenAI GPT-5.6 Sol, OpenAI GPT-5.5, Anthropic Claude Opus 4.8 and Google Gemma 4 meet our language-support criteria in Khmer. Only the solutions on the first two meet general translation quality.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Anthropic Claude Sonnet 5 | 0.766 | 9 | 1.33 | 9 | No | Not shortlisted |
| DeepSeek V3.2 | 0.764 | 11 | 1.41 | 12 | No | Not shortlisted |
| OpenAI GPT-4.1 | 0.758 | 14 | 1.46 | 14 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.751 | 15 | 1.43 | 13 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.750 | 16 | 1.56 | 15 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.735 | 18 | 1.74 | 17 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.733 | 19 | 1.75 | 18 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.732 | 20 | 2.24 | 21 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.727 | 21 | 2.30 | 22 | No | Not shortlisted |
| Mistral Large 3 | 0.714 | 22 | 2.10 | 20 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.645 | 23 | 3.67 | 23 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.641 | 24 | 4.08 | 24 | No | Not shortlisted |
| Cohere Tiny Aya | 0.590 | 25 | 5.94 | 25 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.521 | 26 | 6.45 | 26 | No | Not shortlisted |
| MiniMax M2.5 | 0.481 | 27 | 8.49 | 27 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.439 | 28 | 9.02 | 28 | No | Not shortlisted |
| Cohere Command A Translate | 0.307 | 29 | 14.72 | 29 | No | Not shortlisted |
| Google Gemini 3.5 Flash | 0.780 | 1 | 1.03 | 1 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.778 | 2 | 1.09 | 2 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | 0.777 | 3 | 1.15 | 3 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.773 | 4 | 1.20 | 5 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.772 | 5 | 1.16 | 4 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.5 | 0.772 | 6 | 1.23 | 6 | Yes | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | 0.770 | 7 | 1.27 | 7 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 4.6 | 0.769 | 8 | 1.37 | 10 | Yes | Top 8 by COMET-Kiwi |
| z.ai GLM-5 | 0.766 | 10 | 1.30 | 8 | Yes | Top 8 by MetricX |
| Google Translation LLM | 0.764 | 12 | 1.64 | 16 | Yes | Real-time place |
| Translated Lara 2 | 0.763 | 13 | 1.39 | 11 | Yes | Real-time place |
| Microsoft Translator | 0.738 | 17 | 1.75 | 19 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| Google Gemini 3.5 Flash | Adjusted LLM | 1 | 0.780 | 1 | 1.03 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 2 | 0.778 | 2 | 1.09 | Top 8 by both metrics |
| OpenAI GPT-5.6 Sol | Adjusted LLM | 3 | 0.777 | 3 | 1.15 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 4 | 0.773 | 5 | 1.20 | Top 8 by both metrics |
| Google Gemma 4 | Adjusted LLM | 5 | 0.772 | 4 | 1.16 | Top 8 by both metrics |
| OpenAI GPT-5.5 | Adjusted LLM | 6 | 0.772 | 6 | 1.23 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 7 | 0.770 | 7 | 1.27 | Top 8 by both metrics |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 8 | 0.769 | 10 | 1.37 | Top 8 by COMET-Kiwi |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 9 | 0.766 | 9 | 1.33 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 10 | 0.766 | 8 | 1.30 | Top 8 by MetricX |
| DeepSeek V3.2 | Adjusted LLM | 11 | 0.764 | 12 | 1.41 | Not shortlisted |
| Google Translation LLM | Adjusted NMT | 12 | 0.764 | 16 | 1.64 | Real-time place |
| Translated Lara 2 | Adjusted NMT | 13 | 0.763 | 11 | 1.39 | Real-time place |
| OpenAI GPT-4.1 | Adjusted LLM | 14 | 0.758 | 14 | 1.46 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 15 | 0.751 | 13 | 1.43 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 16 | 0.750 | 15 | 1.56 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 17 | 0.738 | 19 | 1.75 | Real-time place |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 18 | 0.735 | 17 | 1.74 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 19 | 0.733 | 18 | 1.75 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 20 | 0.732 | 21 | 2.24 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 21 | 0.727 | 22 | 2.30 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 22 | 0.714 | 20 | 2.10 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 23 | 0.645 | 23 | 3.67 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Off-the-shelf LLM | 24 | 0.641 | 24 | 4.08 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 25 | 0.590 | 25 | 5.94 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 26 | 0.521 | 26 | 6.45 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 27 | 0.481 | 27 | 8.49 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 28 | 0.439 | 28 | 9.02 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 29 | 0.307 | 29 | 14.72 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.776 | — | 1.15 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.774 | — | 1.33 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.742 | — | 1.62 | 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 29 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 in Khmer. Translations from Google Gemma 4, 5th on COMET-Kiwi and 4th on MetricX, have 53.4 weighted errors per 1,000 words, against 34.3 for Anthropic Claude Opus 4.8, 7th on both. COMET-Kiwi alone put Anthropic Claude Sonnet 4.6 on the shortlist and MetricX alone z.ai GLM-5, and translations from them have 60.0 and 71.5 weighted errors per 1,000 words. Nearest the cut in the tail is Anthropic Claude Sonnet 5, 9th on both metrics.
| 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 | $12.04 |
| Standard | OpenAI GPT-5.5, one-step LLM adjusted for Standard requirements | $17.75 |
| Advanced | OpenAI GPT-5.6 Sol, two-step with the gender-neutrality agent | $27.43 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.17 | Open-weight LLMs |
| z.ai GLM-5 | $4.19 | LLMs |
| Google Gemini 3.5 Flash | $4.39 | LLMs |
| Google Gemini 3.1 Pro Preview | $5.91 | LLMs |
| Microsoft Translator | $10 | MT engines |
| OpenAI GPT-5.6 Sol | $12.04 | LLMs |
| Anthropic Claude Sonnet 4.6 | $13.81 | LLMs |
| OpenAI GPT-5.5 | $17.75 | LLMs |
| Google Translation LLM | $21.15 | LLM-based MT engines |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| Anthropic Claude Opus 4.8 | $25.65 | LLMs |
| Anthropic Claude Fable 5 | $50.69 | LLMs |
| Model | Tokens per character | Class |
|---|---|---|
| Cohere Command A Translate | 1.518 | LLMs |
| MiniMax M2.5 | 1.308 | LLMs |
| Mistral Large 3 | 0.99 | LLMs |
| Mistral Medium 3.1 | 0.983 | LLMs |
| DeepSeek V3.2 | 0.974 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.96 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 0.956 | LLMs |
| Anthropic Claude Sonnet 5 | 0.899 | LLMs |
| Anthropic Claude Fable 5 | 0.897 | LLMs |
| Anthropic Claude Opus 4.8 | 0.893 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.888 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.842 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.834 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.829 | LLMs |
| z.ai GLM-5 | 0.808 | LLMs |
| OpenAI GPT-5.6 Sol | 0.528 | LLMs |
| OpenAI GPT-5.5 | 0.526 | LLMs |
| OpenAI GPT-5.6 Luna | 0.524 | LLMs |
| OpenAI GPT-5.6 Terra | 0.524 | LLMs |
| OpenAI GPT-4.1 | 0.514 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.401 | LLMs |
| Google Gemini 3.5 Flash | 0.4 | LLMs |
| Google Gemma 4 | 0.399 | Open-weight LLMs |
| Cohere Tiny Aya | 0.373 | LLMs |
With billing on tokens, the tokenizers produce a median 0.84 tokens per character of Khmer, 2.9 times their count for English. On the shortlist, the tokenizers of the Google LLMs produce 0.399 to 0.401 tokens per character, those of the OpenAI models 0.526 and 0.528, and those of the Anthropic models 0.842 to 0.897. Steps add cost too. The one-step solution on OpenAI GPT-5.6 Sol costs $12.04 per million characters, its two-step solution (translation + post-edit) $27.09 and with the gender-neutrality agent $27.43.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | OpenAI GPT-5.5, one-step LLM adjusted for Standard requirements | 216 |
| Standard | OpenAI GPT-5.5, one-step LLM adjusted for Standard requirements | 216 |
| Advanced | OpenAI GPT-5.6 Sol, two-step with the gender-neutrality agent | 432 |
| 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 | 0.4 | MT engines |
| Google Translation LLM | 0.7 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 3.7 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 4.3 | MT engines |
| Translated Lara 2 | 6.3 | LLM-based MT engines |
| Cohere Tiny Aya | 19 | LLMs |
| OpenAI GPT-5.6 Luna | 21 | LLMs |
| Google Gemini 3.5 Flash | 22 | LLMs |
| Alibaba Machine Translation API | 31 | MT engines |
| OpenAI GPT-4.1 | 46 | LLMs |
| DeepSeek V3.2 | 57 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 74 | LLMs |
| OpenAI GPT-5.6 Terra | 78 | LLMs |
| Google Gemini 3.1 Pro Preview | 95 | LLMs |
| Cohere Command A Translate | 104 | LLMs |
| Anthropic Claude Haiku 4.5 | 114 | LLMs |
| Google Gemma 4 | 154 | Open-weight LLMs |
| Anthropic Claude Sonnet 4.6 | 163 | LLMs |
| xAI Grok 4.20 (reasoning) | 180 | LLMs |
| Mistral Medium 3.1 | 192 | LLMs |
| MiniMax M2.5 | 204 | LLMs |
| OpenAI GPT-5.5 | 216 | LLMs |
| OpenAI GPT-5.6 Sol | 253 | LLMs |
| Anthropic Claude Sonnet 5 | 270 | LLMs |
| Mistral Large 3 | 302 | LLMs |
| Anthropic Claude Fable 5 | 341 | LLMs |
| Anthropic Claude Opus 4.8 | 361 | LLMs |
| z.ai GLM-5 | 363 | LLMs |
| Alibaba Qwen3-235B-A22B | 371 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 439 | 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 Khmer.
Five of the six real-time models in Khmer are MT engines or LLM-based MT engines. Translations from the three on the shortlist, Google Translation LLM, Translated Lara 2 and Microsoft Translator, have 83.4, 85.7 and 127.9 weighted errors per 1,000 words, with 9, 2 and 3 critical errors. The sixth, the open-weight OpenAI gpt-oss-120b, is the fastest LLM, at 3.7 seconds per 1,000 words, and is off the shortlist.
| 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 |
|---|---|---|---|---|---|
| OpenAI GPT-5.6 Sol | two-step, with the gender-neutrality agent | 100% | ✓ | $27.43 | 432 |
| OpenAI GPT-5.5 | three-step, with the gender-neutrality agent | 100% | ✓ | $1,021 | 475 |
For tone of voice in Khmer we check the formal register, with honorific vocabulary. For gender neutrality we check that a translation does not add a gender where the source leaves it open. Khmer does not have grammatical gender for persons, so the task is to avoid gendered terms. Without the gender-neutrality agent, the solutions on OpenAI GPT-5.5 and OpenAI GPT-5.6 Sol have 99.1% of the sentences about people gender-neutral, and with it both of the two best solutions have 100%.
The three LQA judges agree on a median 93.6% of the sentences of the shortlisted solutions in Khmer. Three of the 12 solutions, on Anthropic Claude Fable 5, Google Gemini 3.1 Pro Preview and Google Gemini 3.5 Flash, do not have a conclusive Intento LQA result, so the general translation quality results on this page come from the solutions on the other nine models. With both pairs of automatic metrics in agreement, the shortlist rests on measurement we can rely on.
Leave feedback