GPT-6 Astra Solves the Last Problem, FrontierMath Tier 4 Officially Saturated
Epoch AI announced that with GPT-6 Astra solving the last previously unsolved problem, all research-level math problems in FrontierMath Tier 4 have been solved by AI, officially saturating the benchmark roughly 14 months after its release. The final problem was authored by mathematician Jay Pantone. This marks another milestone in the rapid climb of frontier models' mathematical reasoning abilities.
Confirmed
- All FrontierMath Tier 4 problems have been solved, officially confirmed by Epoch AI. According to @量子位, the problem solved by GPT-6 Astra was the only one never previously cracked by AI.
- The final problem's author, mathematician Jay Pantone, noted the set was designed in the era of o4-mini, with difficulty calibrated to models at that time.
- OpenAI reportedly posted a Tier 4 score of 97.6%; Claude Fable 5 also achieved a result, though the exact figure is incomplete in the sources.
Unconfirmed
- The exact Claude Fable 5 score cited by @量子位 cannot be fully verified from the available material.
Why it matters
- Mathematicians reviewing AI solutions had often found models using "unexpected shortcuts" on Tier 4 problems—correct answers via non-intended paths—raising questions about the benchmark's validity.
- @Afinetheorem compared earlier forecasts (e.g., LEAP expert predictions from summer 2025), noting saturation came far faster than expected.
- @DotCSV framed it as another landmark in the rapid rise of frontier math reasoning: the hardest ceiling was broken again, and the benchmark retires with its mission accomplished.
2026-09-11 ~ 2026-09-11 · 7 related posts
Primary sources
- GPT-6 Astra solves final FrontierMath Tier 4 problem, completing the benchmark — Miles_Brundage · 2026-09-11
- [source] FrontierMath Tier 4 fully solved 14 months after launch; LEAP panel forecasts lag reality — Afinetheorem · 2026-09-11
- FrontierMath Tier 4, once the hardest math benchmark, falls after 1.5 years — Jsevillamol · 2026-09-11
4 near-duplicate retellings: NathanpmYoung · soumitrashukla9 · Jsevillamol · 量子位