PSEBench: leaderboard
Metric: Verdict accuracy (%, M1): three-way triage accuracy (Reportable, Non-Reportable, Uncertain) over all cases, parse failures counted wrong, on the 5,074-case MN29 benchmark (3,455 complete, 1,362 missing-information and 257 uncertain cases built from clause cards of the Minnesota 29 reportable adverse health events), agentic environment where the model may ASK a GPT-5.2 information provider within 10 turns and then answers with a verdict, clause, legal evidence and rationale; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 95.3 |
| 2 | GPT-5.5 | 94.6 |
| 3 | Claude Opus 4.7 | 92.3 |
| 4 | GPT-5 | 90.2 |
| 5 | Gemini 2.5 Flash | 85.8 |
| 6 | Claude Sonnet 4.6 | 84.5 |
| 7 | DeepSeek R1 | 83.7 |
| 8 | GPT-OSS-120B | 79 |
| 9 | Qwen 3 235B A22B 2507 Instruct | 76.7 |
| 10 | GPT-5 Nano | 71.9 |
| 11 | Mistral Small 3.2 | 60.5 |
| 12 | Llama 3.1 8B Instruct | 52.4 |
Interactive version: theaggregate.ai/benchmark?slug=psebench · How It Works · Data refreshed daily, snapshot 2026-09-29.