The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Punjabi 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 Punjabi. |
| Tone of voice | The formal register (ਤੁਸੀਂ) 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 Punjabi, 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 10 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 Punjabi. Optimizing for cost or latency selects the two-step solution (translation + post-edit) on OpenAI GPT-5.6 Sol, with 29.7 total errors per 1,000 words, at $10.88 per million characters and 258 seconds per 1,000 words. Optimizing for quality selects the three-step solution (translation + LQA + repair) on OpenAI GPT-5.5, with 26.7 total errors per 1,000 words, at $599 per million characters and 594 seconds per 1,000 words.
Under a cost cap from $10.88 to under $594 per million characters, or a time limit from 258 to under 352 seconds per 1,000 words, only the two-step solution (translation + post-edit) on OpenAI GPT-5.6 Sol meets every Standard requirement in Punjabi. Next on both cost and time is the three-step solution (translation + LQA + repair) on Google Gemini 3.1 Pro Preview, with 35.3 total errors per 1,000 words. With one step only, the Basic requirements are the most a solution meets, through the one on OpenAI GPT-5.6 Sol.
For the Advanced requirements, the gender-neutrality agent works with the two best solutions. With the three-step solution (translation + LQA + repair) on OpenAI GPT-5.5, it raises the cost from $599 to $1,000 per million characters and the time from 594 to 679 seconds per 1,000 words. Total errors fall from 26.7 to 17.5 per 1,000 words, and weighted errors rise from 13.5 to 17.5 per 1,000 words. With the agent, the solution on OpenAI GPT-5.6 Sol costs $980 per million characters and takes 391 seconds per 1,000 words.
OpenAI GPT-5.5
three-step
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Sol
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
three-step
OpenAI GPT-5.6 Sol
two-step
OpenAI GPT-5.6 Sol
two-step
OpenAI GPT-5.5
three-step with the gender-neutrality agent
OpenAI GPT-5.5
three-step with the gender-neutrality agent
OpenAI GPT-5.5
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 | 30.1 | $10.88 | 258 |
| Standard | Tier 2: two or more solutions, none simple | ☆☆☆ | two-step | OpenAI GPT-5.6 Sol | 29.7 | $10.88 | 258 |
| Advanced | Tier 3: one solution | ☆ | with the gender-neutrality agent | OpenAI GPT-5.5 | 17.5 | $1,000 | 679 |
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 step up from Basic to Standard keeps the same three models in Punjabi, but the solution on OpenAI GPT-5.6 Sol takes two steps (translation + post-edit), as its one-step version has 99.6% of the tags in place. At both levels, optimizing for quality selects the three-step solution (translation + LQA + repair) on OpenAI GPT-5.5, at 26.7 total errors per 1,000 words, and for cost or latency the solution on OpenAI GPT-5.6 Sol, at 30.1 in one step under the Basic requirements and 29.7 in two under the Standard ones.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.2 | 69.9 | Misses a Standard requirement |
| Mistral Large 3 | 2.94 | 71.5 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 5.12 | 49.1 | Misses a Standard requirement |
| z.ai GLM-5 | 6.21 | 54.7 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 6.73 | 38.2 | Misses a Standard requirement |
| OpenAI GPT-5.6 Sol | 10.88 | 16.8 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 15.58 | 30.7 | Misses a Standard requirement |
| Google Translation LLM | 21.36 | 109.4 | Misses a Standard requirement |
| Translated Lara 2 | 24.99 | 67.9 | Misses a Standard requirement |
| DeepL Next Gen | 27.5 | 54.1 | Misses a Standard requirement |
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 54.4 |
| Gender neutrality | 14.9 |
| Tag handling | 1.2 |
| Tone of voice | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Punjabi | Class |
|---|---|---|
| Google Translation LLM | 109.4 | LLM-based MT engines |
| Mistral Large 3 | 71.5 | LLMs |
| Google Gemma 4 | 69.9 | Open-weight LLMs |
| Translated Lara 2 | 67.9 | LLM-based MT engines |
| z.ai GLM-5 | 54.7 | LLMs |
| DeepL Next Gen | 54.1 | LLM-based MT engines |
| Google Gemini 3.5 Flash | 49.1 | LLMs |
| Google Gemini 3.1 Pro Preview | 38.2 | LLMs |
| OpenAI GPT-5.5 | 30.7 | LLMs |
| OpenAI GPT-5.6 Sol | 16.8 | LLMs |
No translation into Punjabi: AI21 Jamba Large 1.7 (cut off part-way, garbage or near-empty, wrong language, wrong script).
The three shortlisted MT engines in Punjabi are all LLM-based. Translations from DeepL Next Gen, Translated Lara 2 and Google Translation LLM have 54.1, 67.9 and 109.4 weighted errors per 1,000 words in one step. Those from Google Translation LLM have 8 critical errors, the only critical errors among the shortlisted one-step solutions. The two shortlisted LLMs outside our language-support criteria are Google Gemma 4, an open-weight LLM, and Mistral Large 3. Translations from them have 69.9 and 71.5 weighted errors per 1,000 words.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| xAI Grok 4.20 (reasoning) | 0.727 | 9 | 1.48 | 9 | No | Not shortlisted |
| DeepSeek V3.2 | 0.726 | 11 | 1.54 | 11 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.721 | 13 | 1.68 | 14 | No | Not shortlisted |
| OpenAI GPT-4.1 | 0.719 | 14 | 1.51 | 10 | No | Not shortlisted |
| Microsoft Translator | 0.714 | 15 | 1.78 | 16 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.708 | 16 | 1.77 | 15 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.706 | 17 | 1.80 | 17 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.705 | 18 | 1.98 | 20 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.705 | 19 | 1.82 | 18 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.702 | 20 | 2.43 | 23 | No | Not shortlisted |
| MiniMax M2.5 | 0.702 | 21 | 1.82 | 19 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.702 | 22 | 1.61 | 12 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.694 | 23 | 2.17 | 22 | No | Not shortlisted |
| Amazon Translate | 0.679 | 24 | 2.43 | 24 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.669 | 25 | 2.62 | 25 | No | Not shortlisted |
| Cohere Tiny Aya | 0.640 | 26 | 3.03 | 26 | No | Not shortlisted |
| AppTek NMT | 0.528 | 27 | 6.87 | 27 | No | Not shortlisted |
| Cohere Command A Translate | 0.526 | 28 | 7.94 | 28 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.504 | 29 | 8.63 | 31 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.504 | 30 | 8.57 | 30 | No | Not shortlisted |
| Anthropic Claude Opus 4.8 | 0.497 | 31 | 8.85 | 32 | No | Not shortlisted |
| Anthropic Claude Sonnet 5 | 0.495 | 32 | 8.85 | 33 | No | Not shortlisted |
| Anthropic Claude Fable 5 | 0.495 | 33 | 8.96 | 34 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.453 | 34 | 8.21 | 29 | No | Not shortlisted |
| OpenAI GPT-5.6 Sol | 0.738 | 1 | 1.27 | 2 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.737 | 2 | 1.28 | 3 | Yes | Top 8 by both metrics |
| DeepL Next Gen | 0.736 | 3 | 1.30 | 4 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.735 | 4 | 1.27 | 1 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.5 | 0.734 | 5 | 1.45 | 8 | Yes | Top 8 by both metrics |
| Google Gemini 3.5 Flash | 0.732 | 6 | 1.30 | 5 | Yes | Top 8 by both metrics |
| z.ai GLM-5 | 0.729 | 7 | 1.36 | 7 | Yes | Top 8 by both metrics |
| Mistral Large 3 | 0.728 | 8 | 1.32 | 6 | Yes | Top 8 by both metrics |
| Google Translation LLM | 0.726 | 10 | 2.03 | 21 | Yes | Real-time place |
| Translated Lara 2 | 0.723 | 12 | 1.66 | 13 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| OpenAI GPT-5.6 Sol | Adjusted LLM | 1 | 0.738 | 2 | 1.27 | Top 8 by both metrics |
| Google Gemma 4 | Adjusted LLM | 2 | 0.737 | 3 | 1.28 | Top 8 by both metrics |
| DeepL Next Gen | Off-the-shelf NMT | 3 | 0.736 | 4 | 1.30 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 4 | 0.735 | 1 | 1.27 | Top 8 by both metrics |
| OpenAI GPT-5.5 | Adjusted LLM | 5 | 0.734 | 8 | 1.45 | Top 8 by both metrics |
| Google Gemini 3.5 Flash | Adjusted LLM | 6 | 0.732 | 5 | 1.30 | Top 8 by both metrics |
| z.ai GLM-5 | Adjusted LLM | 7 | 0.729 | 7 | 1.36 | Top 8 by both metrics |
| Mistral Large 3 | Adjusted LLM | 8 | 0.728 | 6 | 1.32 | Top 8 by both metrics |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 9 | 0.727 | 9 | 1.48 | Not shortlisted |
| Google Translation LLM | Adjusted NMT | 10 | 0.726 | 21 | 2.03 | Real-time place |
| DeepSeek V3.2 | Adjusted LLM | 11 | 0.726 | 11 | 1.54 | Not shortlisted |
| Translated Lara 2 | Adjusted NMT | 12 | 0.723 | 13 | 1.66 | Real-time place |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 13 | 0.721 | 14 | 1.68 | Not shortlisted |
| OpenAI GPT-4.1 | Adjusted LLM | 14 | 0.719 | 10 | 1.51 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 15 | 0.714 | 16 | 1.78 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 16 | 0.708 | 15 | 1.77 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 17 | 0.706 | 17 | 1.80 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 18 | 0.705 | 20 | 1.98 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Off-the-shelf LLM | 19 | 0.705 | 18 | 1.82 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 20 | 0.702 | 23 | 2.43 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 21 | 0.702 | 19 | 1.82 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 22 | 0.702 | 12 | 1.61 | Not shortlisted |
| SYSTRAN PNMT API | Off-the-shelf NMT | 23 | 0.694 | 22 | 2.17 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 24 | 0.679 | 24 | 2.43 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 25 | 0.669 | 25 | 2.62 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 26 | 0.640 | 26 | 3.03 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 27 | 0.528 | 27 | 6.87 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 28 | 0.526 | 28 | 7.94 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 29 | 0.504 | 31 | 8.63 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 30 | 0.504 | 30 | 8.57 | Not shortlisted |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 31 | 0.497 | 32 | 8.85 | Not shortlisted |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 32 | 0.495 | 33 | 8.85 | Not shortlisted |
| Anthropic Claude Fable 5 | Adjusted LLM | 33 | 0.495 | 34 | 8.96 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 34 | 0.453 | 29 | 8.21 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.737 | — | 1.27 | Same model family as OpenAI GPT-5.6 Sol |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.735 | — | 1.39 | Same model family as OpenAI GPT-5.6 Sol |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.723 | — | 1.53 | 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 34 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.
COMET-Kiwi ranks Google Gemma 4 second in Punjabi and MetricX third. By Intento LQA, translations from the one-step solution on Google Gemma 4 have 69.9 weighted errors per 1,000 words, against 30.7 for the one on OpenAI GPT-5.5, fifth on COMET-Kiwi and eighth on MetricX. Translations from DeepL Next Gen, third on COMET-Kiwi and fourth on MetricX, have 54.1 weighted errors per 1,000 words. Nearest the cut in the tail is xAI Grok 4.20 (reasoning), ninth on both COMET-Kiwi and 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 | $10.88 |
| Standard | OpenAI GPT-5.6 Sol, two-step | $10.88 |
| Advanced | OpenAI GPT-5.5, three-step with the gender-neutrality agent | $1,000 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.2 | Open-weight LLMs |
| Mistral Large 3 | $2.94 | LLMs |
| Google Gemini 3.5 Flash | $5.12 | LLMs |
| z.ai GLM-5 | $6.21 | LLMs |
| Google Gemini 3.1 Pro Preview | $6.73 | LLMs |
| OpenAI GPT-5.6 Sol | $10.88 | LLMs |
| OpenAI GPT-5.5 | $15.58 | LLMs |
| Google Translation LLM | $21.36 | LLM-based MT engines |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Model | Tokens per character | Class |
|---|---|---|
| Mistral Medium 3.1 | 1.857 | LLMs |
| Mistral Large 3 | 1.856 | LLMs |
| Anthropic Claude Opus 4.8 | 1.613 | LLMs |
| Anthropic Claude Fable 5 | 1.61 | LLMs |
| Anthropic Claude Sonnet 4.6 | 1.61 | LLMs |
| Anthropic Claude Sonnet 5 | 1.609 | LLMs |
| Anthropic Claude Haiku 4.5 | 1.579 | LLMs |
| z.ai GLM-5 | 1.297 | LLMs |
| MiniMax M2.5 | 1.284 | LLMs |
| Alibaba Qwen3-235B-A22B | 1.214 | Open-weight LLMs |
| Cohere Command A Translate | 1.079 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.787 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.773 | LLMs |
| DeepSeek V3.2 | 0.744 | LLMs |
| MoonshotAI Kimi K2.5 | 0.64 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.49 | LLMs |
| Google Gemma 4 | 0.488 | Open-weight LLMs |
| Google Gemini 3.5 Flash | 0.485 | LLMs |
| OpenAI GPT-4.1 | 0.472 | LLMs |
| OpenAI GPT-5.6 Luna | 0.47 | LLMs |
| OpenAI GPT-5.6 Sol | 0.469 | LLMs |
| OpenAI GPT-5.6 Terra | 0.467 | LLMs |
| OpenAI GPT-5.5 | 0.465 | LLMs |
| Cohere Tiny Aya | 0.411 | LLMs |
The tokenizers of the two OpenAI models on the shortlist produce 0.465 and 0.469 tokens per character of the translation into Punjabi, and that of Mistral Large 3 produces 1.856. Even so, the one-step solution on Mistral Large 3 costs $2.94 per million characters, less than the $10.88 of the one on OpenAI GPT-5.6 Sol. On OpenAI GPT-5.5, the cost follows the steps: $15.58 per million characters in one step, $47.02 in two (translation + post-edit) and $599 in three (translation + LQA + repair).
| 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 | 258 |
| Standard | OpenAI GPT-5.6 Sol, two-step | 258 |
| Advanced | OpenAI GPT-5.5, three-step with the gender-neutrality agent | 679 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.2 | MT engines |
| ModernMT Enterprise Edition | 0.4 | MT engines |
| Google Translation LLM | 0.9 | LLM-based MT engines |
| DeepL Next Gen | 1.1 | LLM-based MT engines |
| Amazon Translate | 1.2 | MT engines |
| Google Cloud Advanced Translation API | 1.3 | MT engines |
| SYSTRAN PNMT API | 2.1 | MT engines |
| Translated Lara 2 | 2.6 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 3.4 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 4.2 | MT engines |
| OpenAI GPT-5.6 Luna | 20 | LLMs |
| Cohere Tiny Aya | 22 | LLMs |
| Google Gemini 3.5 Flash | 25 | LLMs |
| Alibaba Machine Translation API | 31 | MT engines |
| AppTek NMT | 41 | MT engines |
| xAI Grok 4.20 (non-reasoning) | 46 | LLMs |
| OpenAI GPT-4.1 | 51 | LLMs |
| DeepSeek V3.2 | 58 | LLMs |
| Swiss AI Apertus-8B-2509 | 66 | Open-weight LLMs |
| Mistral Medium 3.1 | 70 | LLMs |
| Anthropic Claude Haiku 4.5 | 74 | LLMs |
| Cohere Command A Translate | 76 | LLMs |
| OpenAI GPT-5.6 Terra | 80 | LLMs |
| Mistral Large 3 | 92 | LLMs |
| Anthropic Claude Sonnet 5 | 105 | LLMs |
| xAI Grok 4.20 (reasoning) | 107 | LLMs |
| Anthropic Claude Fable 5 | 114 | LLMs |
| MiniMax M2.5 | 123 | LLMs |
| Anthropic Claude Opus 4.8 | 132 | LLMs |
| Google Gemini 3.1 Pro Preview | 135 | LLMs |
| Anthropic Claude Sonnet 4.6 | 161 | LLMs |
| Google Gemma 4 | 182 | Open-weight LLMs |
| OpenAI GPT-5.5 | 230 | LLMs |
| OpenAI GPT-5.6 Sol | 258 | LLMs |
| MoonshotAI Kimi K2.5 | 271 | LLMs |
| Alibaba Qwen3-235B-A22B | 482 | Open-weight LLMs |
| z.ai GLM-5 | 765 | 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 Punjabi.
Of the nine real-time (<2 sec) models in Punjabi, five are MT engines, three LLM-based MT engines and one the open-weight OpenAI gpt-oss-120b. On the shortlist, the real-time solutions are the three on the LLM-based MT engines DeepL Next Gen, Google Translation LLM and Translated Lara 2. None of the three meet general translation quality, tag handling or formatting. Under the Advanced requirements, the time rises to 679 seconds per 1,000 words.
| 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-5.5 | three-step, with the gender-neutrality agent | 100% | ✓ | $1,000 | 679 |
| OpenAI GPT-5.6 Sol | three-step, with the gender-neutrality agent | 98.1% | ✗ | $980 | 391 |
For Punjabi we check the formal register, with ਤੁਸੀਂ as the form of address, and gender-neutral wording through neutral nouns and neutral rephrasing, with no default gender. With the gender-neutrality agent, the solution on OpenAI GPT-5.6 Sol has 9.2 weighted errors per 1,000 words, against 17.5 for the one on OpenAI GPT-5.5. Its translations have 99.8% of the tags in place, where tag handling asks for 100%, and it misses that requirement as well as gender neutrality.
For the 10 shortlisted solutions in Punjabi, the median share of sentences on which the three LQA judges agree is 95.5%. None of them has its Intento LQA result limited by disagreement among the judges. Both pairs of automatic metrics agree above their thresholds, so we can rank the solutions in Punjabi on COMET-Kiwi and MetricX.
Leave feedback