AcuityBench (Conversational): leaderboard

AcuityBench (2026) tests whether language models recommend the right urgency of care (four acuity levels) for user medical presentations; in this conversational format the model answers naturally and a GPT-4.1 judge infers the level it communicated. Exact-match accuracy in percent on 527 clear consensus cases; higher is better.

Metric: Exact acuity match (%) in the conversational format on the 527 clear consensus cases: the model answers naturally without being asked for a label, a GPT-4.1 judge infers the communicated acuity level, mode of five samples at temperature 1; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Flash78#463
2GPT-5.477.2#135
3DeepSeek V3.176.8#511
4Gemini 2.5 Pro76.7#276
5Claude Sonnet 4.674.4#173
6Claude Opus 4.771.9#85
7GPT-4.171.4#505
8Claude Haiku 4.5 (20251001)70#543
9GPT-5 Mini67.7#364
10Qwen 2.5 72B Instruct Turbo66.8#1213
11Qwen 2.5 7B Turbo Instruct65.1#2617
12Llama 3.3 70B Instruct61.9#1385

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

Interactive version: theaggregate.ai/benchmark?slug=acuitybench-conversational · How It Works · Data refreshed daily, snapshot 2026-11-21.