NewtonBench - Vanilla Agent: leaderboard
Discovering an altered physical law by experiment: across 324 tasks in 12 physics domains the agent runs experiments on a simulated system whose law has been counterfactually changed, then states the law. Symbolic accuracy, averaged over difficulty levels and four runs, with experiments as the only tool. Higher is better. Source: arxiv.org/abs/2510.07172.
Metric: Symbolic Accuracy (%; 324 tasks, mean of four runs). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 75.9 | #164 |
| 2 | Gemini 2.5 Pro | 65.4 | #263 |
| 3 | GPT-5 Mini | 53.1 | #355 |
| 4 | Gemini 2.5 Flash | 50.6 | #464 |
| 5 | O4 Mini | 47.8 | #363 |
| 6 | DeepSeek R1 | 43.4 | #485 |
| 7 | Qwen 3 235B A22B | 30.9 | #650 |
| 8 | QwQ-32B | 28.5 | #965 |
| 9 | DeepSeek V3 | 8.9 | #745 |
| 10 | GPT-4.1 | 5.8 | #507 |
| 11 | GPT-4.1 Mini | 5.1 | #826 |
No result here: #4 Claude Opus 5.5, #5 GPT-6 Astra, #9 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=newtonbench-vanilla-agent · How It Works · Data refreshed daily, snapshot 2026-11-08.