HAL ScienceAgentBench: leaderboard
Princeton HAL leaderboard for ScienceAgentBench: 102 data-driven scientific discovery tasks from 42 peer-reviewed papers across 4 disciplines.
Metric: Accuracy (%). Source: hal.cs.princeton.edu. Saturation forecast: Around October 2027. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 (Medium) | 33.33 | #124 (O3) |
| 2 | GPT-5 (Medium) | 30.39 | #98 (GPT-5) |
| 3 | Claude Sonnet 4.5 (High) | 30.39 | #142 (Claude Sonnet 4.5) |
| 4 | Claude Sonnet 4.5 | 29.41 | #142 |
| 5 | O4 Mini (High) | 27.45 | #180 (O4 Mini) |
| 6 | Claude Opus 4.1 (20250805) | 27.45 | #132 |
| 7 | O4 Mini (Low) | 27.45 | #180 (O4 Mini) |
| 8 | GPT-4.1 (2025-04-14) | 24.51 | #250 |
| 9 | DeepSeek R1 | 23.53 | #251 |
| 10 | Claude Haiku 4.5 (High) | 23.53 | #280 (Claude Haiku 4.5) |
| 11 | Claude 3.7 Sonnet (20250219) | 22.55 | #201 |
| 12 | Claude Haiku 4.5 (20251001) | 18.63 | #261 |
| 13 | Gemini 2.0 Flash (001) | 12.75 | #360 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=hal-scienceagentbench · How It Works · Data refreshed daily, snapshot 2026-10-20.