NewtonBench - Code Assistance: leaderboard
NewtonBench law discovery with a Python interpreter available alongside the experiments. Same 324 tasks and symbolic-accuracy scoring as the vanilla setting. 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 | 73.3 | #164 |
| 2 | Gemini 2.5 Pro | 63.7 | #263 |
| 3 | GPT-5 Mini | 48.1 | #355 |
| 4 | O4 Mini | 45.1 | #363 |
| 5 | Gemini 2.5 Flash | 40.9 | #464 |
| 6 | DeepSeek R1 | 37.4 | #485 |
| 7 | Qwen 3 235B A22B | 33.7 | #650 |
| 8 | QwQ-32B | 31.4 | #965 |
| 9 | GPT-4.1 | 18.9 | #507 |
| 10 | GPT-4.1 Mini | 14.7 | #826 |
| 11 | DeepSeek V3 | 13.3 | #745 |
No result here: #4 Claude Opus 5.5, #5 GPT-6 Astra, #9 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=newtonbench-code-assistance · How It Works · Data refreshed daily, snapshot 2026-11-08.