GPT-6 Astra Solves Last FrontierMath Tier 4 Problem; Epoch AI Declares Tier Saturated

On September 11, 2026, OpenAI's GPT-6 Astra solved the last previously unsolved problem in Tier 4 of FrontierMath, a research-level mathematics benchmark. OpenAI reported a score of 97.6%, and Epoch AI declared the tier saturated: under its cumulative counting method, every problem in Tier 4 has now been solved at least once by an AI system.

FrontierMath was developed by Epoch AI with more than 60 mathematicians and released on November 7, 2024. It was designed to avoid the rapid saturation that conventional math benchmarks were experiencing. The benchmark initially consisted of 300 original problems across Tiers 1 to 3. Tier 4 was added in 2025 with 50 research-level problems created by mathematics professors and postdoctoral researchers. In June 2026, after an independent audit, Epoch AI revised 12 Tier 4 problems, removed 7, and left 43.

Performance on Tier 4 has climbed quickly: GPT-5.6 Sol reached 83.0%, Claude Fable 5 reached 90.2%, and GPT-6 Astra reached 97.6%, including the one problem that had remained unsolved by any AI. Jay Pantone, a Marquette University associate professor of mathematics who designed problems for the benchmark, said that while AI had previously found numerical shortcuts, this time its solution was quite close to his own approach.

Epoch AI has not concluded that mathematics itself has been solved. FrontierMath now also includes Open Problems, which use unsolved research questions to test models, and FrontierMath Erdős, which requires AI to produce fully formalized proofs. GPT-6 Astra solved only 2 of 68 problems in the Erdős subset.