Language

ThaiExam: cost and speed vs accuracy

111 models, each scored on the same 375 questions (onet_m6, ic, tpat1). Generated 2026-09-17.

jev-latest (TypeSafe) thinking models (45) standard models (65)

Thinking = the run returned reasoning content, or the preset declares it (an effort variant such as [high], a thinking budget, or a reasoning family like o1/o3/r1/qwq).

Cost vs accuracy

Cost per question, log scale. Up and to the left is better.

jev-latest scores 69.1% at $0.00003 per question.

Speed vs accuracy

Mean seconds per question, log scale. Up and to the left is better.

jev-latest scores 69.1% at 0.35s per question.

Original data and results: ai-vs-thai-exams.pages.dev. Exam questions from the ThaiExam dataset (SCB 10X) and the OpenThaiGPT evaluation dataset.