FrontierPhysics v0.1
Science
BenchFlow's FrontierPhysics v0.1 tests whether a coding agent can carry out end-to-end physics research. Its 56 tasks come from research the contributing physicists did themselves, across nine domains (12 condensed matter, 7 AMO, 6 astrophysics, 6 fluid dynamics, 6 statistical, 5 quantum information, 5 photonics and imaging, 5 high-energy and 4 chemical physics). Each task asks for a literature review and plan graded against the author's expert rubric, an implementation checked by code verifiers, and deliverables including a paper. Every configuration runs three trials per task (168 rollouts) in a developer agent harness (Claude Code, Codex, Antigravity or Grok Build) through the BenchFlow runtime, and rollouts that were never evaluated count as failures. The headline Overall Pass, which the leaderboard ranks by, is the share of the 168 rollouts that pass. It is stricter than passing the code verifier alone, but the page does not publish its exact rule.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 22.6 |
| GPT-6 Astra | 17.9 |
| GPT-6.1 Sol | 16.1 |
| Claude Fable 5.1 | 13.1 |
| GPT-6 Sol | 11.3 |
| Claude Opus 5 | 10.1 |
| Gemini 4 Argon | 8.3 |
| GPT-5.6 Sol | 8.3 |
| Grok 4.7 | 6.5 |
| Claude Haiku 5.5 | 4.2 |
| GPT-5.6 Luna | 3.6 |
| GPT-5.6 Terra | 3.0 |
| Grok 4.6 | 3.0 |
| GLM-5.3 | 1.2 |
| MiMo-V2.6-Pro | 1.2 |
| Kimi K3 | 0.6 |
| Mistral Large 4 Preview | 0.6 |
| Gemini 3.8 Flash | 0.0 |