GPT-6 closed the last problem on AI's hardest math test. A harder one is already live.

FrontierMath is a math benchmark built by Epoch AI and introduced in late 2024 as a deliberate foil to AI. Epoch hired roughly 60 mathematicians, including Fields medalists Terence Tao, Timothy Gowers, and Richard Borcherds, to write research-grade problems drawn from their own work. When the project added a fourth tier in July 2025, explicitly built to resist AI for years, Tao called it "极其困难" (extremely difficult). The top score on Tier 4 that month was about 5%.

Fourteen months later, that tier is fully cleared. OpenAI's GPT-6 Astra hit 97.6% on Epoch's v2 audit of the 43-problem tier and closed the one previously-untouched problem. Epoch AI officially declared the tier "saturated." The score is not 100% for any single model. "Saturated" means every problem has been solved at least once across models and attempts, with progression running GPT-5.6 Sol at 83.0%, Claude Fable 5 at 90.2%, and Astra at 97.6%.

The audit was triggered by OpenAI flagging more errors than Epoch expected. Epoch ran GPT-5.5 and Claude Opus 4.7 to triage, then human mathematicians to verify the fixes. Twelve problems were fixed, seven removed, leaving 43.

Jay Pantone, a Marquette University combinatorialist, said Astra's solution on one problem tracked his own approach and that he is no longer surprised by Astra's pace.

FrontierMath has not stopped. Epoch has added "Open Problems," unsolved research questions, and a Lean-formalized Erdős tier, where Astra solved 2 of 68.