AstaBench: leaderboard
Research agents, each driven here by one model in a named agent such as Ai2's ReAct baseline or Smolagents, work through 11 scientific-research tasks (literature search and synthesis, research-code execution, data-driven discovery, end-to-end experiments); the score is the mean of the four category averages, kept only for agents that ran every task. Unlike single-skill science benchmarks it rates the whole agent system across the research workflow, so the agent and its tools vary between rows.
Metric: Mean of the four category scores, agents that ran all 11 test-set tasks (0-1). Source: allenai-asta-bench-leaderboard.hf.space. Saturation forecast: Around February 2027. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (xHigh) | 0.58 | #79 (Claude Opus 4.7) |
| 2 | GPT-5.4 (2026-03-05) (xHigh) | 0.46 | #87 (GPT-5.4 (2026-03-05)) |
| 3 | GPT-5 (Medium) | 0.43 | #167 (GPT-5) |
| 4 | O3 (2025-04-16) (Medium) | 0.38 | #216 (O3 (2025-04-16)) |
| 5 | Claude Sonnet 4 (20250514) (Thinking) | 0.37 | #387 (Claude Sonnet 4 (20250514)) |
| 6 | GPT-5 Mini (2025-08-07) (Medium) | 0.31 | #343 (GPT-5 Mini (2025-08-07)) |
| 7 | GPT-4.1 | 0.3 | #495 |
| 8 | Claude 3.5 Haiku (20241022) | 0.21 | #1578 |
| 9 | GPT-4o (2024-08-06) | 0.16 | #740 |
| 10 | Gemini 2.5 Flash (Preview 05-20) | 0.15 | #426 |
| 11 | Llama 4 Scout Instruct | 0.08 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-13.