We shortlist 10 to 13 solutions per language, including at least three real-time solutions.
We combine two top-8 lists and add real-time solutions. For each language we take the top-8 solutions by COMET-Kiwi and the top-8 by MetricX. A model that runs in several harness variants, or a family of sibling models with statistically indistinguishable scores, counts once, so each list holds one entry per model. The two lists together are the candidates: where the metrics agree the lists overlap and the shortlist is short, where they disagree it is longer. We then look at how many candidates are real-time solutions, a measured median of two seconds or less per document, and if there are fewer than three we add the next real-time solutions by metric rank until there are three, up to 13 solutions in total.
The Afrikaans shortlist contains 13 solutions. the two top-8 lists share four solutions, so the candidates are twelve; two of them are real-time solutions, so one more is added, and the shortlist has 13.
Shortlisting solutions for Afrikaans
| Step | Solutions | Which |
|---|---|---|
| 1. In the COMET-Kiwi top-8 | 8 | OpenAI GPT-5.6 Terra, Anthropic Claude Haiku 4.5, Grok 4.20 (reasoning), DeepL Next Gen (0.55 s), Google Gemma 4, Anthropic Claude Sonnet 5, Translated Lara 2 (1.14 s), Anthropic Claude Fable 5 |
| 2. In the MetricX top-8 | 8 | OpenAI GPT-5.6 Terra, Grok 4.20 (reasoning), DeepL Next Gen (0.55 s), Google Gemma 4, OpenAI GPT-5.5, Anthropic Claude Sonnet 4.6, Google Gemini 3.5 Flash, MoonshotAI Kimi K2.5 |
| 3. In both top-8 lists | 4 | OpenAI GPT-5.6 Terra, Grok 4.20 (reasoning), DeepL Next Gen (0.55 s), Google Gemma 4 |
| 4. Candidates: either top-8, each model once | 12 | OpenAI GPT-5.6 Terra, Anthropic Claude Haiku 4.5, Grok 4.20 (reasoning), DeepL Next Gen (0.55 s), Google Gemma 4, Anthropic Claude Sonnet 5, Translated Lara 2 (1.14 s), Anthropic Claude Fable 5, OpenAI GPT-5.5, Anthropic Claude Sonnet 4.6, Google Gemini 3.5 Flash, MoonshotAI Kimi K2.5 |
| 5. Real-time solutions already among them (median 2 s or less per document) | 2 | DeepL Next Gen (0.55 s), Translated Lara 2 (1.14 s) |
| 6. Added to reach three real-time solutions, next by metric rank | 1 | Microsoft Translator (0.17 s) |
| 7. Shortlist | 13 |
Latencies in brackets are the median seconds per document of the real-time solutions.
The shortlist per language, with the metric that selected each entry
| Name | places | both | comet only | metricx only | real time |
|---|---|---|---|---|---|
| Afrikaans | 13 | 4 | 4 | 4 | 1 |
| Albanian | 11 | 5 | 3 | 3 | 0 |
| Amharic | 11 | 7 | 1 | 1 | 2 |
| Arabic (MSA) | 13 | 5 | 3 | 3 | 2 |
| Armenian | 11 | 7 | 1 | 1 | 2 |
| Assamese | 11 | 0 | 10 | 0 | 1 |
| Azerbaijani | 11 | 7 | 1 | 1 | 2 |
| Basque | 12 | 6 | 2 | 2 | 2 |
| Belarusian | 11 | 6 | 2 | 2 | 1 |
| Bengali | 13 | 4 | 4 | 4 | 1 |
| Bosnian | 11 | 5 | 3 | 3 | 0 |
| Bulgarian | 12 | 6 | 2 | 2 | 2 |
| Burmese | 11 | 7 | 1 | 1 | 2 |
| Catalan | 13 | 4 | 4 | 4 | 1 |
| Cebuano | 12 | 0 | 10 | 0 | 2 |
| Chinese Simplified | 12 | 6 | 2 | 2 | 2 |
| Chinese Traditional | 13 | 3 | 5 | 5 | 0 |
| Croatian | 13 | 4 | 4 | 4 | 1 |
| Czech | 12 | 6 | 2 | 2 | 2 |
| Danish | 12 | 6 | 2 | 2 | 2 |
| Dutch | 13 | 4 | 4 | 4 | 1 |
| Estonian | 12 | 6 | 2 | 2 | 2 |
| Farsi | 12 | 6 | 2 | 2 | 2 |
| Finnish | 11 | 6 | 2 | 2 | 1 |
| French | 13 | 5 | 3 | 3 | 2 |
| Frisian | 13 | 5 | 3 | 3 | 2 |
| Galician | 12 | 5 | 3 | 3 | 1 |
| Georgian | 11 | 7 | 1 | 1 | 2 |
| German | 11 | 5 | 3 | 3 | 0 |
| Greek | 11 | 7 | 1 | 1 | 2 |
| Gujarati | 13 | 3 | 5 | 5 | 0 |
| Hausa | 11 | 6 | 2 | 2 | 1 |
| Hebrew | 12 | 6 | 2 | 2 | 2 |
| Hindi | 13 | 5 | 3 | 3 | 2 |
| Hungarian | 12 | 6 | 2 | 2 | 2 |
| Icelandic | 13 | 5 | 3 | 3 | 2 |
| Indonesian | 12 | 6 | 2 | 2 | 2 |
| Irish | 12 | 5 | 3 | 3 | 1 |
| Italian | 12 | 4 | 4 | 4 | 0 |
| Japanese | 13 | 4 | 4 | 4 | 1 |
| Javanese | 13 | 3 | 5 | 5 | 0 |
| Kannada | 12 | 7 | 1 | 1 | 3 |
| Kazakh | 12 | 6 | 2 | 2 | 2 |
| Khmer | 12 | 7 | 1 | 1 | 3 |
| Kinyarwanda | 11 | 0 | 10 | 0 | 1 |
| Korean | 12 | 6 | 2 | 2 | 2 |
| Kurdish (Kurmanji) | 12 | 6 | 2 | 2 | 2 |
| Kyrgyz | 11 | 6 | 2 | 2 | 1 |
| Lao | 11 | 7 | 1 | 1 | 2 |
| Latvian | 11 | 7 | 1 | 1 | 2 |
| Lithuanian | 12 | 6 | 2 | 2 | 2 |
| Macedonian | 11 | 6 | 2 | 2 | 1 |
| Malagasy | 12 | 6 | 2 | 2 | 2 |
| Malay | 11 | 6 | 2 | 2 | 1 |
| Malayalam | 10 | 8 | 0 | 0 | 2 |
| Marathi | 13 | 5 | 3 | 3 | 2 |
| Mongolian | 10 | 8 | 0 | 0 | 2 |
| Nepali | 11 | 7 | 1 | 1 | 2 |
| Norwegian | 12 | 6 | 2 | 2 | 2 |
| Odia (Oriya) | 11 | 0 | 10 | 0 | 1 |
| Pashto | 10 | 7 | 1 | 1 | 1 |
| Polish | 12 | 5 | 3 | 3 | 1 |
| Portuguese (European) | 11 | 7 | 1 | 1 | 2 |
| Punjabi | 10 | 8 | 0 | 0 | 2 |
| Romanian | 12 | 6 | 2 | 2 | 2 |
| Russian | 13 | 4 | 4 | 4 | 1 |
| Scottish Gaelic | 12 | 7 | 1 | 1 | 3 |
| Serbian | 13 | 4 | 4 | 4 | 1 |
| Shona | 11 | 0 | 10 | 0 | 1 |
| Sinhala | 12 | 7 | 1 | 1 | 3 |
| Slovak | 11 | 7 | 1 | 1 | 2 |
| Slovenian | 13 | 4 | 4 | 4 | 1 |
| Somali | 11 | 7 | 1 | 1 | 2 |
| Spanish (European) | 11 | 5 | 3 | 3 | 0 |
| Sundanese | 12 | 5 | 3 | 3 | 1 |
| Swahili | 11 | 7 | 1 | 1 | 2 |
| Swedish | 11 | 7 | 1 | 1 | 2 |
| Tagalog | 12 | 6 | 2 | 2 | 2 |
| Tajik | 12 | 0 | 10 | 0 | 2 |
| Tamil | 12 | 6 | 2 | 2 | 2 |
| Tatar | 11 | 0 | 10 | 0 | 1 |
| Telugu | 12 | 5 | 3 | 3 | 1 |
| Thai | 13 | 6 | 2 | 2 | 3 |
| Turkish | 13 | 5 | 3 | 3 | 2 |
| Ukrainian | 13 | 5 | 3 | 3 | 2 |
| Urdu | 12 | 6 | 2 | 2 | 2 |
| Uzbek | 11 | 8 | 0 | 0 | 3 |
| Vietnamese | 12 | 6 | 2 | 2 | 2 |
| Welsh | 11 | 6 | 2 | 2 | 1 |
| Wolof | 12 | 0 | 10 | 0 | 2 |
| Xhosa | 13 | 3 | 5 | 5 | 0 |
| Yiddish | 12 | 6 | 2 | 2 | 2 |
| Yoruba | 12 | 0 | 10 | 0 | 2 |
| Zulu | 10 | 0 | 10 | 0 | 0 |
Units — places: places; both: places; comet only: places; metricx only: places; real time: places
'40 solutions' in the report sources counts distinct models on the shortlist; model × configuration pairs are 53.