PatientAgentBench: leaderboard
PatientAgentBench (2026) evaluates patient-facing health AI agents: a simulated patient converses for up to 15 turns with the model inside a baseline agent harness with 15 healthcare sandbox tools, and a jury of Claude Opus 4.8 and GPT-5.5 scores each of 1,200 conversations on six clinician-grounded rubrics. The board is the weighted aggregate rubric score from 1 to 5, with clinical safety weighted most; higher is better.
Metric: Weighted aggregate score (1-5): mean rubric score weighted 2.0 clinical safety, 1.6 workflow accuracy, 1.4 triage quality, 1.4 clinical helpfulness, 1.0 task completion and 0.9 conversational quality; 1,200 synthetic patient conversations of up to 15 turns with a simulated patient, the model inside the benchmark's baseline ReAct agent harness with 15 healthcare sandbox tools; jury of Claude Opus 4.8 and GPT-5.5 judges scoring clinician-grounded rubrics on a 1 to 5 scale; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 4.25 | #56 |
| 2 | GPT-5.5 | 4.22 | #49 |
| 3 | Claude Sonnet 5 | 4.2 | #111 |
| 4 | GPT-5.4 | 4.16 | #123 |
| 5 | Gemini 3 Flash | 3.78 | #164 |
| 6 | Gemini 3.1 Pro (Preview) | 3.71 | #94 |
| 7 | Claude Haiku 4.5 | 3.63 | #568 |
| 8 | GPT-OSS-120B | 3.45 | #700 |
| 9 | Qwen 3 235B A22B | 3.4 | #623 |
| 10 | Qwen3-Next-80B | 3.1 | #626 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=patientagentbench · How It Works · Data refreshed daily, snapshot 2026-11-13.