The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Norwegian 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 Norwegian. |
| Tone of voice | The formal register (du, where direct address cannot be avoided) 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 Norwegian, 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. 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.
Five one-step solutions on LLMs meet every Standard requirement in Norwegian. Optimize for selects OpenAI GPT-5.5 for quality, with 9.9 total errors per 1,000 words at $10.11 per million characters and 50 seconds per 1,000 words. For cost and latency it selects OpenAI GPT-5.6 Luna, with 17.1 total errors per 1,000 words at $0.40 per million characters and 12 seconds per 1,000 words. Those on Google Gemini 3.1 Pro Preview and Anthropic Claude Opus 4.8 have more errors than the one on OpenAI GPT-5.6 Luna, cost more and take longer.
Under the Standard requirements, a cost cap of at most $4.30 per million characters or a time limit below 50 seconds per 1,000 words leaves at most one solution, the one on OpenAI GPT-5.6 Luna. A limit of one step leaves all five, as each of them is a one-step solution. On the shortlist, the real-time (<2 sec) solutions on DeepL Next Gen and Microsoft Translator miss general translation quality, tag handling and formatting, and the one on OpenAI gpt-oss-120b misses general translation quality.
The Advanced requirements add gender neutrality, counted on the two best solutions, those on OpenAI GPT-5.5 and Anthropic Claude Fable 5. Without the gender-neutrality agent, 100% of the sentences about people in their translations are gender-neutral. Under the Advanced requirements, the cheapest solution costs $10.11 per million characters instead of $0.40, and the fastest takes 50 seconds per 1,000 words instead of 12. The solution on OpenAI GPT-5.5 still has the fewest, with 9.9 total errors per 1,000 words.
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Luna
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Luna
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Luna
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Luna
one-step LLM adjusted for Standard requirements
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
one-step LLM adjusted for Standard requirements
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 | ★×5 | one-step | OpenAI GPT-5.6 Luna | 17.1 | $0.40 | 12 |
| Standard | Tier 1: simple solution + backup | ★×5 | one-step | OpenAI GPT-5.6 Luna | 17.1 | $0.40 | 12 |
| Advanced | Tier 1: simple solution + backup | ★★ | one-step | OpenAI GPT-5.5 | 9.9 | $10.11 | 50 |
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 same five one-step solutions meet the Basic and the Standard requirements, as all five also meet tone of voice, tag handling and formatting. For the Advanced requirements, only the two best solutions, those on OpenAI GPT-5.5 and Anthropic Claude Fable 5, count for gender neutrality. Every level takes one step, without the gender-neutrality agent. Optimize for selects OpenAI GPT-5.6 Luna for cost and latency under the Basic and Standard requirements, and OpenAI GPT-5.5 in every other case.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.13 | 25 | Misses a Standard requirement |
| OpenAI GPT-5.6 Luna | 0.4 | 17.1 | Meets every Standard requirement |
| Mistral Large 3 | 0.73 | 31.6 | Misses a Standard requirement |
| OpenAI GPT-4.1 | 2.81 | 31.6 | Misses a Standard requirement |
| Google Gemini 3.1 Pro Preview | 4.31 | 17.8 | Meets every Standard requirement |
| Anthropic Claude Sonnet 5 | 5.46 | 30.3 | Misses a Standard requirement |
| Microsoft Translator | 10 | 65.3 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 10.11 | 9.9 | Meets every Standard requirement |
| Anthropic Claude Opus 4.8 | 13.81 | 18.5 | Meets every Standard requirement |
| Anthropic Claude Fable 5 | 27.4 | 10.9 | Meets every Standard requirement |
| DeepL Next Gen | 27.5 | 21.4 | Misses a Standard requirement |
Among the solutions that meet every Standard requirement, the front of cost and errors holds the two on OpenAI GPT-5.6 Luna and OpenAI GPT-5.5. Across the 11 shortlisted one-step solutions with a cost, the cheapest, on Google Gemma 4 at $0.13 per million characters, has 25.0 weighted errors per 1,000 words. The dearest, on DeepL Next Gen at $27.50 per million characters, has 21.4 weighted errors per 1,000 words. The one on Microsoft Translator costs $10.00 per million characters and has 65.3 weighted errors per 1,000 words.
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 23.2 |
| Tone of voice | 0 |
| Tag handling | 0 |
| Gender neutrality | 0 |
The requirements with countable errors. Every solution with its result on each requirement: the explorer.
Fewer shortlisted solutions meet general translation quality than tone of voice, tag handling or formatting, and the five that do are all on LLMs. Translations from Google Gemma 4 and OpenAI GPT-4.1 have three major errors each, within the bar, but 25.0 and 31.6 weighted errors per 1,000 words. Translations from DeepL Next Gen and Microsoft Translator have 93.3% and 96.5% of the source tags, and those from every shortlisted LLM 100%. With one-step solutions alone, Norwegian is in tier 1 (simple solution + backup).
| Model | Weighted errors per 1,000 words of the shortlisted one-step solutions in Norwegian | Class |
|---|---|---|
| Microsoft Translator | 65.3 | MT engines |
| OpenAI gpt-oss-120b | 52.1 | Open-weight LLMs |
| OpenAI GPT-4.1 | 31.6 | LLMs |
| Mistral Large 3 | 31.6 | LLMs |
| Anthropic Claude Sonnet 5 | 30.3 | LLMs |
| Google Gemma 4 | 25 | Open-weight LLMs |
| DeepL Next Gen | 21.4 | LLM-based MT engines |
| Anthropic Claude Opus 4.8 | 18.5 | LLMs |
| Google Gemini 3.1 Pro Preview | 17.8 | LLMs |
| OpenAI GPT-5.6 Luna | 17.1 | LLMs |
| Anthropic Claude Fable 5 | 10.9 | LLMs |
| OpenAI GPT-5.5 | 9.9 | LLMs |
No translation into Norwegian: Widn AI Tower Anthill (cut off part-way, wrong language), Widn AI Tower Vesuvius (cut off part-way).
The five shortlisted solutions with the fewest weighted errors per 1,000 words are all on LLMs. On the shortlist, the eight LLM solutions have 9.9 to 31.6 weighted errors per 1,000 words, the two open-weight LLMs 25.0 and 52.1, DeepL Next Gen, the LLM-based MT engine, 21.4, and Microsoft Translator, the MT engine, 65.3. All 10 shortlisted LLMs, open-weight ones included, support Norwegian by our measured criteria, and 5 of them meet general translation quality.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Google Gemini 3.5 Flash | 0.812 | 10 | 1.27 | 17 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.811 | 12 | 1.14 | 13 | No | Not shortlisted |
| DeepSeek V3.2 | 0.810 | 13 | 1.11 | 9 | No | Not shortlisted |
| z.ai GLM-5 | 0.809 | 14 | 1.16 | 14 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.809 | 15 | 1.23 | 16 | No | Not shortlisted |
| xAI Grok 4.20 (reasoning) | 0.808 | 16 | 1.12 | 11 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.808 | 17 | 1.18 | 15 | No | Not shortlisted |
| Cohere Command A Translate | 0.807 | 18 | 1.27 | 18 | No | Not shortlisted |
| Supertext LLM | 0.805 | 20 | 1.47 | 26 | No | Not shortlisted |
| Google Translation LLM | 0.803 | 22 | 1.55 | 29 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.803 | 23 | 1.31 | 21 | No | Not shortlisted |
| Tilde TildeOpen LLM engine | 0.803 | 24 | 1.38 | 24 | No | Not shortlisted |
| SpeakLeash Bielik-11B-v3.0-Instruct | 0.803 | 25 | 1.50 | 28 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.801 | 26 | 1.35 | 23 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.798 | 27 | 1.33 | 22 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.795 | 28 | 1.63 | 31 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.795 | 29 | 1.39 | 25 | No | Not shortlisted |
| Widn AI Tower Sugarloaf | 0.795 | 30 | 1.47 | 27 | No | Not shortlisted |
| MiniMax M2.5 | 0.790 | 31 | 1.71 | 32 | No | Not shortlisted |
| NiuTrans | 0.786 | 32 | 1.58 | 30 | No | Not shortlisted |
| Amazon Translate | 0.783 | 33 | 1.76 | 33 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.779 | 34 | 1.85 | 34 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.774 | 35 | 2.09 | 36 | No | Not shortlisted |
| Oracle Cloud Language Service | 0.770 | 36 | 1.89 | 35 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.757 | 37 | 2.27 | 37 | No | Not shortlisted |
| Cohere Tiny Aya | 0.747 | 38 | 2.42 | 38 | No | Not shortlisted |
| AppTek NMT | 0.724 | 39 | 3.05 | 39 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.678 | 40 | 3.36 | 40 | No | Not shortlisted |
| DeepL Next Gen | 0.819 | 1 | 1.06 | 5 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.6 Luna | 0.818 | 2 | 0.95 | 2 | Yes | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | 0.817 | 3 | 0.94 | 1 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.817 | 4 | 1.08 | 7 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | 0.816 | 5 | 1.11 | 10 | Yes | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.5 | 0.816 | 6 | 1.00 | 3 | Yes | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | 0.816 | 7 | 1.10 | 8 | Yes | Top 8 by both metrics |
| Anthropic Claude Fable 5 | 0.814 | 8 | 1.13 | 12 | Yes | Top 8 by COMET-Kiwi |
| Mistral Large 3 | 0.813 | 9 | 1.01 | 4 | Yes | Top 8 by MetricX |
| OpenAI GPT-4.1 | 0.812 | 11 | 1.06 | 6 | Yes | Top 8 by MetricX |
| OpenAI gpt-oss-120b | 0.807 | 19 | 1.27 | 19 | Yes | Real-time place |
| Microsoft Translator | 0.804 | 21 | 1.28 | 20 | Yes | Real-time place |
| Model | Solution complexity | COMET-Kiwi rank | COMET-Kiwi | MetricX rank | MetricX | Shortlist |
|---|---|---|---|---|---|---|
| DeepL Next Gen | Off-the-shelf NMT | 1 | 0.819 | 5 | 1.06 | Top 8 by both metrics |
| OpenAI GPT-5.6 Luna | Adjusted LLM | 2 | 0.818 | 2 | 0.95 | Top 8 by both metrics |
| Google Gemini 3.1 Pro Preview | Adjusted LLM | 3 | 0.817 | 1 | 0.94 | Top 8 by both metrics |
| Google Gemma 4 | Adjusted LLM | 4 | 0.817 | 7 | 1.08 | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 5 | 0.816 | 10 | 1.11 | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.5 | Adjusted LLM | 6 | 0.816 | 3 | 1.00 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 7 | 0.816 | 8 | 1.10 | Top 8 by both metrics |
| Anthropic Claude Fable 5 | Adjusted LLM | 8 | 0.814 | 12 | 1.13 | Top 8 by COMET-Kiwi |
| Mistral Large 3 | Adjusted LLM | 9 | 0.813 | 4 | 1.01 | Top 8 by MetricX |
| Google Gemini 3.5 Flash | Adjusted LLM | 10 | 0.812 | 17 | 1.27 | Not shortlisted |
| OpenAI GPT-4.1 | Adjusted LLM | 11 | 0.812 | 6 | 1.06 | Top 8 by MetricX |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 12 | 0.811 | 13 | 1.14 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 13 | 0.810 | 9 | 1.11 | Not shortlisted |
| z.ai GLM-5 | Adjusted LLM | 14 | 0.809 | 14 | 1.16 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 15 | 0.809 | 16 | 1.23 | Not shortlisted |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | 16 | 0.808 | 11 | 1.12 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 17 | 0.808 | 15 | 1.18 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 18 | 0.807 | 18 | 1.27 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 19 | 0.807 | 19 | 1.27 | Real-time place |
| Supertext LLM | Off-the-shelf NMT | 20 | 0.805 | 26 | 1.47 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 21 | 0.804 | 20 | 1.28 | Real-time place |
| Google Translation LLM | Adjusted NMT | 22 | 0.803 | 29 | 1.55 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Adjusted LLM | 23 | 0.803 | 21 | 1.31 | Not shortlisted |
| Tilde TildeOpen LLM engine | Off-the-shelf NMT | 24 | 0.803 | 24 | 1.38 | Not shortlisted |
| SpeakLeash Bielik-11B-v3.0-Instruct | Adjusted LLM | 25 | 0.803 | 28 | 1.50 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 26 | 0.801 | 23 | 1.35 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 27 | 0.798 | 22 | 1.33 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 28 | 0.795 | 31 | 1.63 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 29 | 0.795 | 25 | 1.39 | Not shortlisted |
| Widn AI Tower Sugarloaf | Adjusted NMT | 30 | 0.795 | 27 | 1.47 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 31 | 0.790 | 32 | 1.71 | Not shortlisted |
| NiuTrans | Off-the-shelf NMT | 32 | 0.786 | 30 | 1.58 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 33 | 0.783 | 33 | 1.76 | Not shortlisted |
| SYSTRAN PNMT API | Off-the-shelf NMT | 34 | 0.779 | 34 | 1.85 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 35 | 0.774 | 36 | 2.09 | Not shortlisted |
| Oracle Cloud Language Service | Off-the-shelf NMT | 36 | 0.770 | 35 | 1.89 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 37 | 0.757 | 37 | 2.27 | Not shortlisted |
| Cohere Tiny Aya | Off-the-shelf LLM | 38 | 0.747 | 38 | 2.42 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 39 | 0.724 | 39 | 3.05 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 40 | 0.678 | 40 | 3.36 | Not shortlisted |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.816 | — | 0.98 | Same model family as OpenAI GPT-5.6 Luna |
| OpenAI GPT-5.6 Terra | Adjusted LLM | — | 0.815 | — | 1.03 | Same model family as OpenAI GPT-5.6 Luna |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | — | 0.802 | — | 1.24 | 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 40 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 shortlist differently. Translations from DeepL Next Gen, first on COMET-Kiwi, have 21.4 weighted errors per 1,000 words, over the bar. OpenAI GPT-5.5, first with 9.9 weighted errors per 1,000 words, is 6th on COMET-Kiwi and 3rd on MetricX. Anthropic Claude Fable 5, second with 10.9 weighted errors per 1,000 words, is on the list through COMET-Kiwi alone, 8th there and 12th on MetricX. The highest-ranked models left out are Google Gemini 3.5 Flash on COMET-Kiwi and DeepSeek V3.2 on MetricX.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.6 Luna, one-step LLM adjusted for Standard requirements | $0.40 |
| Standard | OpenAI GPT-5.6 Luna, one-step LLM adjusted for Standard requirements | $0.40 |
| Advanced | OpenAI GPT-5.5, one-step LLM adjusted for Standard requirements | $10.11 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.13 | Open-weight LLMs |
| OpenAI GPT-5.6 Luna | $0.4 | LLMs |
| Mistral Large 3 | $0.73 | LLMs |
| OpenAI GPT-4.1 | $2.81 | LLMs |
| Google Gemini 3.1 Pro Preview | $4.31 | LLMs |
| Anthropic Claude Sonnet 5 | $5.46 | LLMs |
| Microsoft Translator | $10 | MT engines |
| OpenAI GPT-5.5 | $10.11 | LLMs |
| Anthropic Claude Opus 4.8 | $13.81 | LLMs |
| Anthropic Claude Fable 5 | $27.4 | LLMs |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Model | Tokens per character | Class |
|---|---|---|
| Anthropic Claude Opus 4.8 | 0.457 | LLMs |
| Anthropic Claude Fable 5 | 0.456 | LLMs |
| Anthropic Claude Sonnet 5 | 0.455 | LLMs |
| Mistral Medium 3.1 | 0.382 | LLMs |
| Mistral Large 3 | 0.38 | LLMs |
| AI21 Jamba Large 1.7 | 0.373 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.346 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.344 | LLMs |
| Cohere Command A Translate | 0.34 | LLMs |
| MoonshotAI Kimi K2.5 | 0.339 | LLMs |
| DeepSeek V3.2 | 0.331 | LLMs |
| Widn AI Tower Sugarloaf | 0.329 | LLM-based MT engines |
| Alibaba Qwen3-235B-A22B | 0.328 | Open-weight LLMs |
| z.ai GLM-5 | 0.314 | LLMs |
| MiniMax M2.5 | 0.302 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.3 | LLMs |
| Google Gemma 4 | 0.299 | Open-weight LLMs |
| Google Gemini 3.5 Flash | 0.298 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.29 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.288 | LLMs |
| OpenAI GPT-5.5 | 0.285 | LLMs |
| OpenAI GPT-4.1 | 0.284 | LLMs |
| OpenAI GPT-5.6 Sol | 0.284 | LLMs |
| OpenAI GPT-5.6 Terra | 0.284 | LLMs |
| OpenAI GPT-5.6 Luna | 0.283 | LLMs |
| Cohere Tiny Aya | 0.262 | LLMs |
Under the Standard requirements, the cheapest solution is the one-step solution on OpenAI GPT-5.6 Luna, and under the Advanced requirements the one-step solution on OpenAI GPT-5.5, both without the gender-neutrality agent. So the higher cost of the Advanced requirements comes with the model, not with steps.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | OpenAI GPT-5.6 Luna, one-step LLM adjusted for Standard requirements | 12 |
| Standard | OpenAI GPT-5.6 Luna, one-step LLM adjusted for Standard requirements | 12 |
| Advanced | OpenAI GPT-5.5, one-step LLM adjusted for Standard requirements | 50 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.3 | MT engines |
| ModernMT Enterprise Edition | 0.4 | MT engines |
| Google Cloud Advanced Translation API | 0.4 | MT engines |
| Google Translation LLM | 1 | LLM-based MT engines |
| Amazon Translate | 1.1 | MT engines |
| DeepL Next Gen | 1.8 | LLM-based MT engines |
| SYSTRAN PNMT API | 1.8 | MT engines |
| OpenAI gpt-oss-120b | 2.4 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 4.1 | MT engines |
| Oracle Cloud Language Service | 4.3 | MT engines |
| OpenAI GPT-5.6 Luna | 12 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 14 | LLMs |
| OpenAI GPT-4.1 | 15 | LLMs |
| Supertext LLM | 15 | LLM-based MT engines |
| Cohere Tiny Aya | 18 | LLMs |
| OpenAI GPT-5.6 Terra | 23 | LLMs |
| Alibaba Machine Translation API | 25 | MT engines |
| Anthropic Claude Haiku 4.5 | 29 | LLMs |
| Mistral Medium 3.1 | 37 | LLMs |
| AppTek NMT | 38 | MT engines |
| MiniMax M2.5 | 39 | LLMs |
| DeepSeek V3.2 | 42 | LLMs |
| Anthropic Claude Sonnet 5 | 44 | LLMs |
| xAI Grok 4.20 (reasoning) | 44 | LLMs |
| Cohere Command A Translate | 46 | LLMs |
| OpenAI GPT-5.6 Sol | 47 | LLMs |
| Swiss AI Apertus-8B-2509 | 48 | Open-weight LLMs |
| Widn AI Tower Sugarloaf | 48 | LLM-based MT engines |
| OpenAI GPT-5.5 | 50 | LLMs |
| Anthropic Claude Fable 5 | 53 | LLMs |
| AI21 Jamba Large 1.7 | 53 | LLMs |
| Mistral Large 3 | 54 | LLMs |
| NiuTrans | 54 | MT engines |
| Anthropic Claude Sonnet 4.6 | 55 | LLMs |
| SpeakLeash Bielik-11B-v3.0-Instruct | 55 | Open-weight LLMs |
| Anthropic Claude Opus 4.8 | 58 | LLMs |
| Google Gemini 3.5 Flash | 59 | LLMs |
| Google Gemini 3.1 Pro Preview | 85 | LLMs |
| Alibaba Qwen3-235B-A22B | 113 | Open-weight LLMs |
| Google Gemma 4 | 130 | Open-weight LLMs |
| z.ai GLM-5 | 158 | LLMs |
| Tilde TildeOpen LLM engine | 183 | LLM-based MT engines |
| MoonshotAI Kimi K2.5 | 242 | 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 Norwegian.
Seven of the eight real-time models in Norwegian are MT engines or LLM-based MT engines, and the other one is the open-weight OpenAI gpt-oss-120b. The solutions that meet every Standard requirement take from 12 seconds per 1,000 words, on OpenAI GPT-5.6 Luna, to 85, on Google Gemini 3.1 Pro Preview. Under the Advanced requirements, the fastest solution takes 50 seconds per 1,000 words, on OpenAI GPT-5.5.
| 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 | one-step | 100% | ✓ | $10.11 | 50 |
| Anthropic Claude Fable 5 | one-step | 100% | ✓ | $27.40 | 53 |
The tone of voice we check in Norwegian is the formal register, with du where direct address cannot be avoided. Translations from Google Gemma 4 have 97.9% of the sentences in that register, and those from the other shortlisted solutions 100%. For gender neutrality we check rephrasing, plural forms and neutral nouns, with hen optional, and a default gender does not meet the requirement. The two best solutions meet it without the gender-neutrality agent, so the cost and time of each stay as under the Standard requirements.
For both pairs of automatic metrics, the whole interval of the agreement is above the threshold: 0.58 to 0.61 for the similarity pair and 0.62 to 0.64 for the quality-estimation pair. The three LQA judges agree on a median of 96.3% of the sentences of the shortlisted solutions. So the results on this page rest on automatic metrics that agree within each pair and on LQA judges that agree on most sentences.
Leave feedback