The State of Translation Automation 2026 · Languages
What it takes to automate translation from English into Macedonian 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 Macedonian. |
| Tone of voice | The formal register (Вие) 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 Macedonian, 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.
Four one-step solutions meet every Standard requirement in Macedonian. The one on OpenAI GPT-5.6 Terra has the fewest total errors, 14.9 per 1,000 words, the lowest cost and the shortest time, so Optimize for selects it for quality, cost and latency alike. Among the other three, the solution on Google Gemini 3.1 Pro Preview costs $5.09 per million characters against $12.01 for the one on OpenAI GPT-5.5, but has 22.5 total errors per 1,000 words against 16.2 and takes 97 seconds per 1,000 words against 73.
Under the Standard requirements, a cost cap or a time limit below those of the solution on OpenAI GPT-5.6 Terra, $4.84 per million characters and 32 seconds per 1,000 words, removes all four solutions. A limit on the steps keeps all four, since each takes one step. Under the Advanced requirements, any limit on the solution complexity removes the only solution, because it works with the gender-neutrality agent. So does a cost cap below $12.57 per million characters or a time limit below 64 seconds per 1,000 words.
The gender-neutrality agent takes the solution on OpenAI GPT-5.5 from $12.01 to $12.57 per million characters, from 73 to 64 seconds per 1,000 words and from 16.2 to 12.2 total errors per 1,000 words. On OpenAI GPT-5.6 Terra, the agent comes in a three-step solution (translation + LQA + repair), so the change combines two extra steps with the agent. That solution has 2.9 total errors per 1,000 words against 14.9 for the one-step solution, at $976 against $4.84 per million characters and 160 against 32 seconds per 1,000 words.
OpenAI GPT-5.6 Terra
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Terra
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Terra
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Terra
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Terra
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.6 Terra
one-step LLM adjusted for Standard requirements
OpenAI GPT-5.5
one-step with the gender-neutrality agent
OpenAI GPT-5.5
one-step with the gender-neutrality agent
OpenAI GPT-5.5
one-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 | ★×4 | one-step | OpenAI GPT-5.6 Terra | 14.9 | $4.84 | 32 |
| Standard | Tier 1: simple solution + backup | ★×4 | one-step | OpenAI GPT-5.6 Terra | 14.9 | $4.84 | 32 |
| Advanced | Tier 3: one solution | ☆ | with the gender-neutrality agent | OpenAI GPT-5.5 | 12.2 | $12.57 | 64 |
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.
Optimize for selects one solution for quality, cost and latency alike: the one-step solution on OpenAI GPT-5.6 Terra under the Basic and Standard requirements, and the one on OpenAI GPT-5.5 with the gender-neutrality agent under the Advanced ones. The step to the Advanced requirements removes all four one-step solutions and takes the agent. The only solution that meets them has 12.2 total errors per 1,000 words against 14.9 for the one on OpenAI GPT-5.6 Terra, at $12.57 per million characters and 64 seconds per 1,000 words.
| Model | US dollars per million characters | Weighted errors per 1,000 words | Group |
|---|---|---|---|
| Google Gemma 4 | 0.15 | 42.8 | Misses a Standard requirement |
| z.ai GLM-5 | 2.1 | 41.9 | Misses a Standard requirement |
| Google Gemini 3.5 Flash | 3.8 | 30.7 | Misses a Standard requirement |
| OpenAI GPT-5.6 Terra | 4.84 | 11.9 | Meets every Standard requirement |
| Google Gemini 3.1 Pro Preview | 5.09 | 19.4 | Meets every Standard requirement |
| Anthropic Claude Sonnet 5 | 5.32 | 48.8 | Misses a Standard requirement |
| Microsoft Translator | 10 | 57.7 | Misses a Standard requirement |
| OpenAI GPT-5.5 | 12.01 | 13.5 | Meets every Standard requirement |
| Translated Lara 2 | 24.99 | 24.7 | Misses a Standard requirement |
| Anthropic Claude Fable 5 | 26.95 | 17.5 | Meets every Standard requirement |
| DeepL Next Gen | 27.5 | 43.2 | Misses a Standard requirement |
| Requirement | Median errors per 1,000 words |
|---|---|
| General translation quality | 30.7 |
| Gender neutrality | 2.8 |
| Tone of voice | 0.1 |
| Tag handling | 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 Macedonian | Class |
|---|---|---|
| Microsoft Translator | 57.7 | MT engines |
| Anthropic Claude Sonnet 5 | 48.8 | LLMs |
| DeepL Next Gen | 43.2 | LLM-based MT engines |
| Google Gemma 4 | 42.8 | Open-weight LLMs |
| z.ai GLM-5 | 41.9 | LLMs |
| Google Gemini 3.5 Flash | 30.7 | LLMs |
| Translated Lara 2 | 24.7 | LLM-based MT engines |
| Google Gemini 3.1 Pro Preview | 19.4 | LLMs |
| Anthropic Claude Fable 5 | 17.5 | LLMs |
| OpenAI GPT-5.5 | 13.5 | LLMs |
| OpenAI GPT-5.6 Terra | 11.9 | LLMs |
On the shortlist, translations from the eight LLMs have 11.9 to 48.8 weighted errors per 1,000 words, and the four solutions that meet general translation quality are all on LLMs. Translations from the two LLM-based MT engines, Translated Lara 2 and DeepL Next Gen, have 24.7 and 43.2 weighted errors per 1,000 words, and translations from the MT engine, Microsoft Translator, have 57.7. All eight shortlisted LLMs, the open-weight Google Gemma 4 among them, support Macedonian by our measured criteria.
| Model | COMET-Kiwi | Rank on COMET-Kiwi | MetricX | Rank on MetricX | Shortlisted | Shortlist place |
|---|---|---|---|---|---|---|
| Anthropic Claude Opus 4.8 | 0.806 | 9 | 1.33 | 10 | No | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | 0.806 | 10 | 1.35 | 14 | No | Not shortlisted |
| OpenAI GPT-4.1 | 0.804 | 13 | 1.34 | 13 | No | Not shortlisted |
| xAI Grok 4.20 (non-reasoning) | 0.802 | 15 | 1.34 | 11 | No | Not shortlisted |
| DeepSeek V3.2 | 0.799 | 16 | 1.44 | 19 | No | Not shortlisted |
| Tilde TildeOpen LLM engine | 0.799 | 17 | 1.39 | 18 | No | Not shortlisted |
| Google Translation LLM | 0.799 | 18 | 1.57 | 24 | No | Not shortlisted |
| Mistral Large 3 | 0.797 | 19 | 1.36 | 15 | No | Not shortlisted |
| ModernMT Enterprise Edition | 0.797 | 20 | 1.37 | 16 | No | Not shortlisted |
| MoonshotAI Kimi K2.5 | 0.794 | 21 | 1.46 | 21 | No | Not shortlisted |
| Mistral Medium 3.1 | 0.791 | 22 | 1.49 | 22 | No | Not shortlisted |
| MiniMax M2.5 | 0.791 | 23 | 1.51 | 23 | No | Not shortlisted |
| Anthropic Claude Haiku 4.5 | 0.790 | 24 | 1.59 | 26 | No | Not shortlisted |
| Swiss AI Apertus-8B-2509 | 0.788 | 25 | 1.61 | 27 | No | Not shortlisted |
| Alibaba Qwen3-235B-A22B | 0.783 | 26 | 1.46 | 20 | No | Not shortlisted |
| SYSTRAN PNMT API | 0.782 | 27 | 1.72 | 28 | No | Not shortlisted |
| OpenAI gpt-oss-120b | 0.778 | 28 | 1.59 | 25 | No | Not shortlisted |
| Amazon Translate | 0.778 | 29 | 1.81 | 29 | No | Not shortlisted |
| Google Cloud Advanced Translation API | 0.771 | 30 | 2.18 | 31 | No | Not shortlisted |
| Cohere Command A Translate | 0.761 | 31 | 1.84 | 30 | No | Not shortlisted |
| Yandex Cloud Translate v2 | 0.742 | 32 | 2.54 | 32 | No | Not shortlisted |
| AI21 Jamba Large 1.7 | 0.693 | 33 | 3.14 | 34 | No | Not shortlisted |
| NiuTrans | 0.687 | 34 | 2.73 | 33 | No | Not shortlisted |
| Alibaba Machine Translation API | 0.624 | 35 | 4.23 | 35 | No | Not shortlisted |
| AppTek NMT | 0.457 | 36 | 8.72 | 36 | No | Not shortlisted |
| Google Gemini 3.1 Pro Preview | 0.812 | 1 | 1.25 | 5 | Yes | Top 8 by both metrics |
| Translated Lara 2 | 0.811 | 2 | 1.39 | 17 | Yes | Top 8 by COMET-Kiwi |
| z.ai GLM-5 | 0.811 | 3 | 1.29 | 9 | Yes | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.6 Terra | 0.811 | 4 | 1.14 | 1 | Yes | Top 8 by both metrics |
| DeepL Next Gen | 0.811 | 5 | 1.27 | 7 | Yes | Top 8 by both metrics |
| Google Gemini 3.5 Flash | 0.808 | 6 | 1.17 | 2 | Yes | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | 0.808 | 7 | 1.23 | 4 | Yes | Top 8 by both metrics |
| Google Gemma 4 | 0.808 | 8 | 1.29 | 8 | Yes | Top 8 by both metrics |
| OpenAI GPT-5.5 | 0.805 | 11 | 1.26 | 6 | Yes | Top 8 by MetricX |
| Anthropic Claude Fable 5 | 0.804 | 12 | 1.22 | 3 | Yes | Top 8 by MetricX |
| Microsoft Translator | 0.804 | 14 | 1.34 | 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 | 0.812 | 5 | 1.25 | Top 8 by both metrics |
| Translated Lara 2 | Adjusted NMT | 2 | 0.811 | 17 | 1.39 | Top 8 by COMET-Kiwi |
| z.ai GLM-5 | Adjusted LLM | 3 | 0.811 | 9 | 1.29 | Top 8 by COMET-Kiwi |
| OpenAI GPT-5.6 Terra | Adjusted LLM | 4 | 0.811 | 1 | 1.14 | Top 8 by both metrics |
| DeepL Next Gen | Off-the-shelf NMT | 5 | 0.811 | 7 | 1.27 | Top 8 by both metrics |
| Google Gemini 3.5 Flash | Adjusted LLM | 6 | 0.808 | 2 | 1.17 | Top 8 by both metrics |
| Anthropic Claude Sonnet 5 | Adjusted LLM | 7 | 0.808 | 4 | 1.23 | Top 8 by both metrics |
| Google Gemma 4 | Adjusted LLM | 8 | 0.808 | 8 | 1.29 | Top 8 by both metrics |
| Anthropic Claude Opus 4.8 | Adjusted LLM | 9 | 0.806 | 10 | 1.33 | Not shortlisted |
| Anthropic Claude Sonnet 4.6 | Adjusted LLM | 10 | 0.806 | 14 | 1.35 | Not shortlisted |
| OpenAI GPT-5.5 | Adjusted LLM | 11 | 0.805 | 6 | 1.26 | Top 8 by MetricX |
| Anthropic Claude Fable 5 | Adjusted LLM | 12 | 0.804 | 3 | 1.22 | Top 8 by MetricX |
| OpenAI GPT-4.1 | Adjusted LLM | 13 | 0.804 | 13 | 1.34 | Not shortlisted |
| Microsoft Translator | Off-the-shelf NMT | 14 | 0.804 | 12 | 1.34 | Real-time place |
| xAI Grok 4.20 (non-reasoning) | Adjusted LLM | 15 | 0.802 | 11 | 1.34 | Not shortlisted |
| DeepSeek V3.2 | Adjusted LLM | 16 | 0.799 | 19 | 1.44 | Not shortlisted |
| Tilde TildeOpen LLM engine | Off-the-shelf NMT | 17 | 0.799 | 18 | 1.39 | Not shortlisted |
| Google Translation LLM | Adjusted NMT | 18 | 0.799 | 24 | 1.57 | Not shortlisted |
| Mistral Large 3 | Adjusted LLM | 19 | 0.797 | 15 | 1.36 | Not shortlisted |
| ModernMT Enterprise Edition | Off-the-shelf NMT | 20 | 0.797 | 16 | 1.37 | Not shortlisted |
| MoonshotAI Kimi K2.5 | Adjusted LLM | 21 | 0.794 | 21 | 1.46 | Not shortlisted |
| Mistral Medium 3.1 | Adjusted LLM | 22 | 0.791 | 22 | 1.49 | Not shortlisted |
| MiniMax M2.5 | Adjusted LLM | 23 | 0.791 | 23 | 1.51 | Not shortlisted |
| Anthropic Claude Haiku 4.5 | Adjusted LLM | 24 | 0.790 | 26 | 1.59 | Not shortlisted |
| Swiss AI Apertus-8B-2509 | Off-the-shelf LLM | 25 | 0.788 | 27 | 1.61 | Not shortlisted |
| Alibaba Qwen3-235B-A22B | Adjusted LLM | 26 | 0.783 | 20 | 1.46 | Not shortlisted |
| SYSTRAN PNMT API | Off-the-shelf NMT | 27 | 0.782 | 28 | 1.72 | Not shortlisted |
| OpenAI gpt-oss-120b | Adjusted LLM | 28 | 0.778 | 25 | 1.59 | Not shortlisted |
| Amazon Translate | Off-the-shelf NMT | 29 | 0.778 | 29 | 1.81 | Not shortlisted |
| Google Cloud Advanced Translation API | Off-the-shelf NMT | 30 | 0.771 | 31 | 2.18 | Not shortlisted |
| Cohere Command A Translate | Adjusted LLM | 31 | 0.761 | 30 | 1.84 | Not shortlisted |
| Yandex Cloud Translate v2 | Off-the-shelf NMT | 32 | 0.742 | 32 | 2.54 | Not shortlisted |
| AI21 Jamba Large 1.7 | Adjusted LLM | 33 | 0.693 | 34 | 3.14 | Not shortlisted |
| NiuTrans | Off-the-shelf NMT | 34 | 0.687 | 33 | 2.73 | Not shortlisted |
| Alibaba Machine Translation API | Off-the-shelf NMT | 35 | 0.624 | 35 | 4.23 | Not shortlisted |
| AppTek NMT | Off-the-shelf NMT | 36 | 0.457 | 36 | 8.72 | Not shortlisted |
| OpenAI GPT-5.6 Luna | Adjusted LLM | — | 0.805 | — | 1.14 | Same model family as OpenAI GPT-5.6 Terra |
| OpenAI GPT-5.6 Sol | Adjusted LLM | — | 0.809 | — | 1.13 | Same model family as OpenAI GPT-5.6 Terra |
| xAI Grok 4.20 (reasoning) | Adjusted LLM | — | 0.795 | — | 1.74 | Same model family as xAI Grok 4.20 (non-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 36 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 Macedonian. COMET-Kiwi alone brought Translated Lara 2 and z.ai GLM-5 onto the list, second and third on it, and translations from them have 24.7 and 41.9 weighted errors per 1,000 words. MetricX alone brought OpenAI GPT-5.5 and Anthropic Claude Fable 5, and the solutions on both meet every Standard requirement. Nearest the cut are Anthropic Claude Opus 4.8, 9th on COMET-Kiwi and 10th on MetricX, and Anthropic Claude Sonnet 4.6, 10th on COMET-Kiwi.
| Quality level | Cheapest solution that meets every requirement of the level | Cost per million characters |
|---|---|---|
| Basic | OpenAI GPT-5.6 Terra, one-step LLM adjusted for Standard requirements | $4.84 |
| Standard | OpenAI GPT-5.6 Terra, one-step LLM adjusted for Standard requirements | $4.84 |
| Advanced | OpenAI GPT-5.5, one-step with the gender-neutrality agent | $12.57 |
| Model | US dollars per million characters | Class |
|---|---|---|
| Google Gemma 4 | $0.15 | Open-weight LLMs |
| z.ai GLM-5 | $2.1 | LLMs |
| Google Gemini 3.5 Flash | $3.8 | LLMs |
| OpenAI GPT-5.6 Terra | $4.84 | LLMs |
| Google Gemini 3.1 Pro Preview | $5.09 | LLMs |
| Anthropic Claude Sonnet 5 | $5.32 | LLMs |
| Microsoft Translator | $10 | MT engines |
| OpenAI GPT-5.5 | $12.01 | LLMs |
| Translated Lara 2 | $24.99 | LLM-based MT engines |
| Anthropic Claude Fable 5 | $26.95 | LLMs |
| DeepL Next Gen | $27.5 | LLM-based MT engines |
| Model | Tokens per character | Class |
|---|---|---|
| AI21 Jamba Large 1.7 | 0.557 | LLMs |
| Mistral Medium 3.1 | 0.43 | LLMs |
| Anthropic Claude Opus 4.8 | 0.426 | LLMs |
| Mistral Large 3 | 0.426 | LLMs |
| Anthropic Claude Fable 5 | 0.425 | LLMs |
| Anthropic Claude Sonnet 5 | 0.425 | LLMs |
| Alibaba Qwen3-235B-A22B | 0.419 | Open-weight LLMs |
| MoonshotAI Kimi K2.5 | 0.408 | LLMs |
| Anthropic Claude Sonnet 4.6 | 0.397 | LLMs |
| Anthropic Claude Haiku 4.5 | 0.394 | LLMs |
| Cohere Command A Translate | 0.376 | LLMs |
| DeepSeek V3.2 | 0.372 | LLMs |
| z.ai GLM-5 | 0.364 | LLMs |
| Google Gemini 3.1 Pro Preview | 0.345 | LLMs |
| Google Gemini 3.5 Flash | 0.345 | LLMs |
| Google Gemma 4 | 0.343 | Open-weight LLMs |
| OpenAI GPT-4.1 | 0.336 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 0.335 | LLMs |
| OpenAI GPT-5.6 Terra | 0.335 | LLMs |
| xAI Grok 4.20 (reasoning) | 0.334 | LLMs |
| OpenAI GPT-5.5 | 0.334 | LLMs |
| OpenAI GPT-5.6 Luna | 0.334 | LLMs |
| MiniMax M2.5 | 0.333 | LLMs |
| OpenAI GPT-5.6 Sol | 0.333 | LLMs |
The tokenizers produce a median 0.37 tokens per character of Macedonian, 1.5 times as many as for English. Tokens alone do not set the price. The tokenizers of OpenAI GPT-5.6 Terra and OpenAI GPT-5.5 produce 0.335 and 0.334 tokens per character, yet the solutions on them cost $4.84 and $12.01 per million characters. The move from one model to the other makes most of the step to the Advanced requirements, since the gender-neutrality agent takes the solution on OpenAI GPT-5.5 only to $12.57 per million characters.
| Quality level | Fastest solution that meets every requirement of the level | Seconds per 1,000 words |
|---|---|---|
| Basic | OpenAI GPT-5.6 Terra, one-step LLM adjusted for Standard requirements | 32 |
| Standard | OpenAI GPT-5.6 Terra, one-step LLM adjusted for Standard requirements | 32 |
| Advanced | OpenAI GPT-5.5, one-step with the gender-neutrality agent | 64 |
| Model | Seconds per 1,000 words | Class |
|---|---|---|
| Microsoft Translator | 0.2 | MT engines |
| Google Cloud Advanced Translation API | 0.2 | MT engines |
| ModernMT Enterprise Edition | 0.4 | MT engines |
| Google Translation LLM | 0.6 | LLM-based MT engines |
| DeepL Next Gen | 0.7 | LLM-based MT engines |
| SYSTRAN PNMT API | 1 | MT engines |
| Amazon Translate | 1.3 | MT engines |
| Translated Lara 2 | 1.5 | LLM-based MT engines |
| OpenAI gpt-oss-120b | 2.5 | Open-weight LLMs |
| Yandex Cloud Translate v2 | 2.8 | MT engines |
| Tilde TildeOpen LLM engine | 6.8 | LLM-based MT engines |
| Google Gemini 3.5 Flash | 16 | LLMs |
| Alibaba Machine Translation API | 16 | MT engines |
| OpenAI GPT-5.6 Luna | 17 | LLMs |
| xAI Grok 4.20 (non-reasoning) | 21 | LLMs |
| OpenAI GPT-4.1 | 22 | LLMs |
| OpenAI GPT-5.6 Terra | 32 | LLMs |
| Swiss AI Apertus-8B-2509 | 41 | Open-weight LLMs |
| Mistral Medium 3.1 | 45 | LLMs |
| MiniMax M2.5 | 45 | LLMs |
| Anthropic Claude Haiku 4.5 | 46 | LLMs |
| xAI Grok 4.20 (reasoning) | 53 | LLMs |
| DeepSeek V3.2 | 58 | LLMs |
| Cohere Command A Translate | 58 | LLMs |
| NiuTrans | 60 | MT engines |
| Mistral Large 3 | 66 | LLMs |
| OpenAI GPT-5.6 Sol | 69 | LLMs |
| OpenAI GPT-5.5 | 73 | LLMs |
| Anthropic Claude Sonnet 5 | 75 | LLMs |
| AI21 Jamba Large 1.7 | 75 | LLMs |
| Anthropic Claude Fable 5 | 79 | LLMs |
| AppTek NMT | 80 | MT engines |
| Anthropic Claude Sonnet 4.6 | 85 | LLMs |
| Google Gemini 3.1 Pro Preview | 97 | LLMs |
| Anthropic Claude Opus 4.8 | 99 | LLMs |
| Alibaba Qwen3-235B-A22B | 157 | Open-weight LLMs |
| Google Gemma 4 | 194 | Open-weight LLMs |
| z.ai GLM-5 | 206 | LLMs |
| MoonshotAI Kimi K2.5 | 250 | 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 Macedonian.
Eight of the nine real-time models in Macedonian are MT engines or LLM-based MT engines. Translations from the three on the shortlist, Microsoft Translator, DeepL Next Gen and Translated Lara 2, have 57.7, 43.2 and 24.7 weighted errors per 1,000 words. None of them meet general translation quality. The solutions that meet every Standard requirement are all on LLMs and take 32 to 97 seconds per 1,000 words, and the one that meets every Advanced requirement takes 64.
| 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 Terra | three-step, with the gender-neutrality agent | 99.1% | ✗ | $976 | 160 |
| OpenAI GPT-5.5 | one-step, with the gender-neutrality agent | 100% | ✓ | $12.57 | 64 |
For tone of voice we check the formal register, with Вие as the form of address. For gender neutrality we check neutral rephrasing with plurals and impersonal forms, and no default gender where the source leaves it open. With the gender-neutrality agent, the three-step solution (translation + LQA + repair) on OpenAI GPT-5.6 Terra has 2.6 weighted errors per 1,000 words, and the one-step solution on OpenAI GPT-5.5 has 12.2. The first costs $976 per million characters and takes 160 seconds per 1,000 words, against $12.57 and 64.
Neither pair of automatic metrics is borderline in Macedonian, so the metric ranks behind the shortlist carry weight. The three LQA judges agree on a median of 95.5% of the sentences across the 11 shortlisted solutions.
Leave feedback