Gamow ScreenBench: leaderboard
ScreenBench — Agents inspect compound-screening plate records to detect and correct hidden errors across182tasks. Mean normalized graded-answer score,0no better than shortcuts and1all answers right. Unprompted condition;the prompt listing possible errors is a separate experiment. Native harness/graded population and score revision retained. Native cohorts differ: historical rows use named cell lines; Astra, Carbon and Fable5.2 use coded names; Hydrogen is unspecified. Sol helped select95tasks; evaluation uses fresh runs.
Metric: Mean normalized answer score (fraction, 0–1). Source: gamowlabs.com. Saturation forecast: Around April 2027. 7 models tracked. Updated 15 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Fable 5.2 | 0.7026 | #1 | |
| 2 | Gemini 4 Hydrogen | Gamow run: native=vanilla_antigravity__gemini-4-hydrogen-eap, access=EAP | 0.7024 | #11 |
| 3 | Gemini 4 Carbon | High | 0.6336 | #2 |
| 4 | Gemini 4 Argon | 0.4923 | #12 | |
| 5 | Claude Opus 5.5 | High | 0.4721 | #5 |
| 6 | GPT-6.1 Sol | High | 0.1243 | #10 |
| 7 | GPT-6 Astra | 0.1169 | #6 |
Interactive version: theaggregate.ai/benchmark?slug=gamow-screenbench · How It Works · Data refreshed daily, snapshot 2026-11-16.