Epoch AI launches FrontierMath Erdős benchmark; GPT-6 Astra first to crack two problems
Jsevillamol · x · 2026-09-04
Epoch AI, with Thomas Bloom of erdosproblems.com, announced FrontierMath Erdős, a formal benchmark of 68 significant unsolved Erdős problems (open as of August 2026), formalized in Lean, requiring AI systems to prove or disprove them within a fixed budget.
Background: Erdős problems have become the de facto yardstick for AI math capability — Bloom's site catalogs 1,217 problems, 652 still unsolved. The new benchmark addresses curation rigor, verification, and replicability.
First result: GPT-6 Astra became the first model to crack problems from the set, solving one on graphs with infinite chromatic number (#74) and one in combinatorial number theory (#126), at a few hundred dollars each.
Related event: Epoch AI Launches FrontierMath Erdős; GPT-6 Astra First to Break Zero(10 posts)→
More from Research
- Three reasons AI is not conscious and may never be, per Nature essay — anilkseth · 2026-10-21
- Aviation tracks every part's provenance — agents can reason over that graph — colinmcnamara · 2026-10-21
- SFT details: instruction/reasoning/code/math mix, synthetic + human traces, no epoching — TimDarcet · 2026-10-21
- Mid-training recipe disclosed: 32k→131k context extension, dropping logit distillation for tokens — TimDarcet · 2026-10-21
- Mathematician Daniel Litt concedes his AI bet, but says AI won't solve math — ziv_ravid · 2026-10-21
- AI drug discovery reality check: 'give me everything that binds,' then filter for toxicity — burny_tech · 2026-10-21