0

ProofBench

Frontier

Automated theorem proving benchmark

Open
Publisher
Vals AI
Domain
Mathematics
Published
Jun 2026
Canonical
vals.ai/benchmarks/proof_bench
Official leaderboard
vals.ai/benchmarks/proof_bench

Cite

Notes

Only stored in your browser.

Attribution

Leaderboard scores
Vals AI
Attribution policy →

Top score 77.0% by Claude Fable 5 - 43 models reporting (25 frontier)

Score history

43
0%25%50%75%100%Aug 25Nov 25Feb 26May 26GPT-5Claude Sonnet 4.5Gemini 3 Pro PreviewClaude Opus 4.6GPT-5.4Claude Opus 4.8Claude Fable 5

Top models

43
ProofBenchBar chart with 21 bars. Highest value: Claude Fable 5 at 77.
21 models

Where it's ranked

1

Official leaderboard

vals.ai

Single benchmark

FAQ

What is ProofBench?
Automated theorem proving benchmark
What is the current top score on ProofBench?
The top reported score is 77.0% by Claude Fable 5, across 43 models reporting (25 from frontier labs).