How many answers were correct
Each bar is one benchmark: how many items it answered correctly out of every item in that set. We wrote the questions and answer keys ourselves and had them checked by two native speakers, and every model answers exactly the same items.
How far the cost pays off
Sorted by the cheapest correct answer first. The left column shows how many items were answered correctly, so a top row with a long left bar means both cheap and accurate.
Tokens used
Input and output are kept apart because they are priced differently, often tenfold. A long output bar means the model talks a lot, and that is where nearly all the cost difference comes from.
Wait time per call
Median wait for a single call on each benchmark. Median rather than mean, so one stalled call does not move the number.
Test coverage
This model has been tested on all 6 benchmarks we publish.
Log9 calls to this model, spread across 1 days.OpenClose
This model is called through a provider id that may not pin a version, so the dates below state which weights were actually tested. Rows marked with a hyphen have a date we inferred from file history rather than recorded at call time, so we do not show a clock time for them.
- 22:35Arti kata Jawa langkaword meaning14/43
- 22:15Arti kata Jawa sehari-hariword meaning98/135
- 21:45Normalisasi teks Indonesiaopen answerPrompt stating the key's rule37/49
- 21:38Normalisasi teks Indonesiaopen answerDetailed prompt forbidding word swaps37/49
- 21:31Normalisasi teks Indonesiaopen answerBuilt-in prompt30/49
- 18:44Koreksi ejaan Indonesiaopen answer26/38
- 08:30Tingkat tutur bahasa Jawaopen answer3/6
- 08:14Tingkat tutur bahasa Jawamultiple choice21/27
- 08:00Pemahaman slang Indonesiamultiple choice40/42