Epoch AI Launches FrontierMath Erdős; GPT-6 Astra First to Break Zero
Epoch AI, together with mathematician and erdosproblems.com founder Thomas Bloom (University of Manchester), released a new math benchmark, FrontierMath Erdős: 68 problems selected from Erdős problems still unsolved as of August 2026, requiring AI systems to produce Lean-formalized, automatically verifiable proofs within a fixed compute budget. GPT-6 Astra became the first public model with a solve rate above 0%, officially solving 2/68 (about 3%); all other top models scored zero.
Confirmed
- The benchmark was built by Epoch AI with Thomas Bloom: 68 unsolved Erdős-type problems answered via Lean-formalized proofs within a fixed compute budget.
- Among five top models tested, GPT-6 Astra was the only nonzero scorer: 2 of 68 problems (about 3%); GPT-5.6 Sol, GPT-5.5, Claude Fable 5.1 and others all scored 0.
- The 3% figure circulating on Reddit and via DrSingularity is consistent with the official 2/68 count.
- A Reddit user posted a chart claiming GPT-6 Astra leads SOL, Fable and Spark on a coding-agent index and ranks first on math benchmarks, but the specific scores in the image could not be verified.
Unconfirmed
- Chinese outlet QbitAI reported Astra "uniquely solved 5" Erdős problems, conflicting with the official 2/68 figure; the official release should be treated as authoritative.
Why it matters
- As @Jsevillamol relayed from Greg Burnham's idea, testing AI on Erdős-level open problems with Lean auto-verification directly probes whether models can produce formal proofs humans have not yet completed, giving a quantifiable yardstick for genuine AI mathematical discovery. Even 2/68 marks the first time a public model breaks zero on such a benchmark.
2026-09-04 ~ 2026-09-05 · 10 related posts
- Episode 1: Epoch AI Launches FrontierMath Erdős; GPT-6 Astra First to Break Zero(2026-09-04, 10 posts)
- Episode 2: GPT-6 Astra Sets Epoch AI ECI Record at 169(2026-09-05, 3 posts)
Primary sources
- Astra becomes first public model to solve any curated hard Erdős problems with Lean proofs — Jsevillamol · 2026-09-04
- Epoch AI launches FrontierMath Erdős benchmark; Astra solves 2 of 68 unsolved problems — littmath · 2026-09-04
- Astra reportedly leads coding agent index over SOL, Fable, and Spark — XCxBigDong69XCx · 2026-09-04
- GPT-6 Astra scores 3% on FrontierMath Erdős benchmark while every other tested model scores 0% — Every_Foundation5197 · 2026-09-05
- GPT-6 Astra reportedly scores 3% on FrontierMath Erdős, spending $220K in compute — haider1 · 2026-09-05
5 near-duplicate retellings: Jsevillamol · keviv9 · 量子位 · Chris-MelodyFirst · haider1