SciAgentBench: leaderboard
Metric: Success rate (%, 0-100) on all 259 tasks; with tools: a ReAct loop over the task-relevant SciAgentGym tools (native function calling, up to 50 tool rounds, 300-second request timeout); SciAgentBench: 259 multi-step scientific tasks (1,134 sub-questions; physics 109, chemistry 81, materials 37, life sciences 32; about 65 percent with images) aggregated from existing benchmarks, kept when four frontier LLMs averaged under 50 percent and SciAgentGym could execute a verified trace; a task counts only when every sub-question is correct (strict JSON matching with 0.05 numeric tolerance, GPT-4.1 checking textual fields); temperature 0.7; higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 41.3 | #91 |
| 2 | Grok 4.1 | 40.3 | #218 |
| 3 | Claude Sonnet 4 | 35.9 | #194 |
| 4 | Gemini 2.5 Flash | 32.7 | #237 |
| 5 | Gemini 2.5 Pro | 32.6 | #145 |
| 6 | O3 | 32 | #121 |
| 7 | O4 Mini | 31.1 | #172 |
| 8 | GLM-4.6V | 30.9 | #309 |
| 9 | Qwen 3 VL 235B A22B (Thinking) | 28 | #228 (Qwen 3 VL 235B A22B) |
| 10 | Qwen 3 VL 32B (Thinking) | 27.9 | #287 (Qwen 3 VL 32B) |
| 11 | Qwen 3 VL 32B Instruct | 27.4 | #276 |
| 12 | Qwen 3 VL 235B A22B Instruct | 23.9 | #264 |
| 13 | Qwen 3 VL 8B Instruct | 23.4 | #401 |
| 14 | Qwen 3 VL 4B Instruct | 19.7 | #506 |
| 15 | GPT-4o | 18.7 | #333 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=sciagentbench · How It Works · Data refreshed daily, snapshot 2026-10-11.