CAM-Bench (Pass@1): leaderboard

CAM-Bench (2026) is a Lean 4 theorem-proving benchmark of 1,000 proof targets adapted from textbook exercises in optimization, numerical linear algebra and numerical analysis. This board is pass@1 in percent, the chance that one sampled whole proof compiles under Lean and Mathlib, estimated from 32 samples; higher is better.

Metric: Pass@1 (%): probability that a single sampled whole proof makes the CAM-Bench Lean file compile under Lean 4 and Mathlib v4.28.0 with integrity checks, the unbiased estimate from 32 samples per target, over the 1,000 proof targets; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)6.04#99
2DeepSeek V4 Pro5.78#196
3Claude Sonnet 4.65.14#173
4GPT-5.42.96#135
5Kimi K2.60.28#192

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=cam-bench-pass-1 · How It Works · Data refreshed daily, snapshot 2026-11-21.