FrontierPhysics - Deterministic Pass@3: leaderboard
FrontierPhysics (You and colleagues, 2026; arXiv:2611.06910v1) measures end-to-end physics research on 56 physicist-authored tasks across nine domains. Agents review literature, plan and run research, then submit numerical outputs and a scientific manuscript. This board reports Deterministic Pass@3: Percentage of 56 tasks with at least one of three trials passing every deterministic verifier. There are three trials per task, 168 designated rollouts per configuration and a 10-hour cap per rollout. Missing outcomes remain failures/zero scores. Each row preserves its native Codex, Claude Code, Grok Build or Antigravity harness and inference effort; exact harness versions and deployment checkpoints are not disclosed. All results use task-set v0.1.
Metric: Percentage of 56 tasks with at least one of three trials passing every deterministic verifier; v0.1, 10-hour cap per rollout; missing outcomes count as zero/failure; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (FrontierPhysics Run: v0.1; harness=Claude Code; effort=max; timeout=10h) | 62.5 | |
| 2 | GPT-6 Astra (FrontierPhysics Run: v0.1; harness=Codex; effort=max; timeout=10h) | 51.8 | |
| 3 | GPT-6 Astra (FrontierPhysics Run: v0.1; harness=Codex; effort=xhigh; timeout=10h) | 51.8 | |
| 4 | GPT-6.1 Sol (FrontierPhysics Run: v0.1; harness=Codex; effort=max; timeout=10h) | 50 | |
| 5 | Gemini 4 Argon (FrontierPhysics Run: v0.1; harness=Antigravity; effort=high; timeout=10h) | 46.4 | |
| 6 | GPT-6 Sol (FrontierPhysics Run: v0.1; harness=Codex; effort=max; timeout=10h) | 44.6 | |
| 7 | Claude Fable 5.1 (FrontierPhysics Run: v0.1; harness=Claude Code; effort=max; timeout=10h) | 44.6 | |
| 8 | GPT-6.1 Terra (FrontierPhysics Run: v0.1; harness=Codex; effort=max; timeout=10h) | 39.3 | |
| 9 | Grok 4.7 (FrontierPhysics Run: v0.1; harness=Grok Build; effort=xhigh; timeout=10h) | 35.7 | |
| 10 | Claude Opus 5 (FrontierPhysics Run: v0.1; harness=Claude Code; effort=max; timeout=10h) | 30.4 |
Interactive version: theaggregate.ai/benchmark?slug=frontierphysics-deterministic-pass-3 · How It Works · Data refreshed daily, snapshot 2026-11-06.