SciAgentBench - L1 (Up to 3 Steps): leaderboard
Metric: Success rate (%, 0-100) on the tasks of difficulty L1 (3 steps or fewer); 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 January 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 58.8 | #91 |
| 2 | Claude Sonnet 4 | 57.4 | #194 |
| 3 | Qwen 3 VL 235B A22B Instruct | 57.1 | #264 |
| 4 | Gemini 2.5 Pro | 54.2 | #145 |
| 5 | O4 Mini | 51 | #172 |
| 6 | Grok 4.1 | 50 | #218 |
| 7 | Qwen 3 VL 4B Instruct | 48.8 | #506 |
| 8 | GLM-4.6V | 48.8 | #309 |
| 9 | Gemini 2.5 Flash | 48 | #237 |
| 10 | O3 | 47.9 | #121 |
| 11 | Qwen 3 VL 32B Instruct | 47.1 | #276 |
| 12 | Qwen 3 VL 235B A22B (Thinking) | 46.8 | #228 (Qwen 3 VL 235B A22B) |
| 13 | Qwen 3 VL 32B (Thinking) | 45.3 | #287 (Qwen 3 VL 32B) |
| 14 | Qwen 3 VL 8B Instruct | 44.4 | #401 |
| 15 | GPT-4o | 36 | #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-l1-up-to-3-steps · How It Works · Data refreshed daily, snapshot 2026-10-11.