AstaBench - End-to-End Discovery: leaderboard
Agents carry a research idea through experiment design, code, runs, analysis and a written report (E2E-Bench and its harder set), graded against task rubrics; the score is the mean of the two. Unlike paper-reproduction benchmarks the agent must design and run a new experiment, and dedicated discovery agents (Asta Panda, Asta CodeScientist) sit beside general harnesses.
Metric: Mean rubric score of E2E-Bench and E2E-Bench-Hard (0-1). Source: allenai-asta-bench-leaderboard.hf.space. Saturation forecast: Around February 2027. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (xHigh) | 0.69 | #79 (Claude Opus 4.7) |
| 2 | Claude Sonnet 4 (20250514) | 0.6 | #387 |
| 3 | Claude 3.7 Sonnet (20250219) | 0.53 | #402 |
| 4 | Claude Sonnet 4 (20250514) (Thinking) | 0.35 | #387 (Claude Sonnet 4 (20250514)) |
| 5 | GPT-5 (Medium) | 0.33 | #167 (GPT-5) |
| 6 | O3 (2025-04-16) (Medium) | 0.21 | #216 (O3 (2025-04-16)) |
| 7 | GPT-5.4 (2026-03-05) (xHigh) | 0.2 | #87 (GPT-5.4 (2026-03-05)) |
| 8 | GPT-4.1 | 0.12 | #495 |
| 9 | GPT-5 Mini (2025-08-07) (Medium) | 0.11 | #343 (GPT-5 Mini (2025-08-07)) |
| 10 | Claude 3.5 Haiku (20241022) | 0.03 | #1578 |
| 11 | GPT-4o (2024-08-06) | 0.01 | #740 |
| 12 | Gemini 2.5 Flash (Preview 05-20) | 0 | #426 |
| 13 | Llama 4 Scout Instruct | 0 | #1543 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=astabench-end-to-end-discovery · How It Works · Data refreshed daily, snapshot 2026-11-13.