Benchmark explainer
What is ProofBench?
Formal theorem-proving capability on graduate-level mathematics problems formalized in Lean 4.
ProofBench as reported in Epoch AI's Capabilities Index CSV.
How it's scored
- Metric
- normalized performance in Epoch CSV, rendered as accuracy (%)
- Score ceiling
- 100
- Construction
- Graduate/advanced-undergraduate problems across analysis, algebra, probability, number theory, logic; v1.1 updated 2026-09-04; binary scoring (proof compiles via Lean 4 checker).
- Human baseline
- No aggregate human baseline; difficulty calibrated to graduate-math level (target: research-grade theorems).
How trustworthy
Discrimination
100/100Does it still separate models?
Saturation headroom
89.9/100How far from ceiling / clustered at the top?
Contamination resistance
14.5/100Public vs held-out; training-leak risk.
Harness comparability
100/100Is it apples-to-apples, or mixed / vendor-optimized?
Freshness
100/100How old is the benchmark?
Contamination history: Public leaderboard with public problem statements; some problems drawn from widely-known results — overlap with training data possible.
Limitations: Lean 4-specific — does not measure proof-construction skill on Coq/Isabelle/Rocq; binary scoring ignores partially-correct proofs; small set, public.
Who leads ProofBench
| Model | Score | Evidence |
|---|---|---|
| Claude Fable 5.1 | 100 | unverified |
| Claude Opus 5.5 | 100 | unverified |
| Claude Sonnet 5.5 | 100 | unverified |
| GPT-6.1 Sol | 99 | unverified |
| Gemini 4 Argon | 99 | unverified |
Compare the top ProofBench scorers
Frequently asked questions
ProofBench: ProofBench as reported in Epoch AI's Capabilities Index CSV.
Claude Fable 5.1 leads ProofBench at 100. The full leaderboard above lists every recorded measurement, not just the headline number.
69 models have recorded scores on ProofBench, spanning a score spread of 100.
tensor.news grades ProofBench B for integrity (score 84/100), ranking #37 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
Yes — top scores are clustering near the ceiling (saturation ratio 0.10), so ProofBench no longer separates leading models well. Treat small gaps at the top with caution.
Its test set is unknown, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on ProofBench are reasonably apples-to-apples.
Every ProofBench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites www.vals.ai/benchmarks/proof_bench.