MINT (Medical) - MedQA (Question First, Initial Answer): leaderboard

MINT (2026) turns medical multiple-choice diagnosis cases into conversations that reveal one clinically labelled evidence shard per turn. In the Ask-Question-First setting the question comes first with an instruction to wait for enough evidence; this board scores the first answer the model commits to on the MedQA (USMLE-style) cases, over the cases it answered. Accuracy in percent; higher is better.

Metric: Initial-answer accuracy (%) on the 600 MedQA-derived diagnosis cases: the multiple-choice question comes first with an instruction to wait, evidence shards follow one per turn, and the first answer the model commits to is scored over the cases it answered (abstentions are left out of the denominator; Ask-Question-First); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.687.7#173
2GPT-5 Mini76.5#364
3O4 Mini76.3#368
4GPT-OSS-20B61.4#1183
5Qwen 3 4B53#2348
6Med42-v2-8B38.7#2558
7Med42-v2-70B38.5
8MINT Qwen2.5-3B (checkpoint unspecified)36.2
9MedGemma 1.5 4B31.8#2716
10MINT OLMo-2-7B (checkpoint unspecified)31.1

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mint-medical-medqa-question-first-initial-answer · How It Works · Data refreshed daily, snapshot 2026-11-21.