Every FrontierMath Tier 4 problem has now been solved by AI

The hardest tier of one of the best-known AI math benchmarks has run out of unsolved problems.

What Epoch reported

Epoch AI says every problem in FrontierMath Tier 4 has now been solved by AI. GPT-6 Astra solved the last problem standing, which was created by mathematician Jay Pantone. Epoch noted that mathematicians had often said AI found unintended shortcuts on Tier 4 problems, and said that was not the case for this final one.

“Every problem solved” is a cumulative claim across attempts rather than a single scored run.

The disclosure that matters

Epoch notes that FrontierMath was developed with funding from OpenAI, which has exclusive access to a subset of the benchmark. That does not change the result, but it is the context readers need when an OpenAI model is the one that closes the set.

Where evaluation goes next

A saturated benchmark stops measuring progress. Epoch has already moved attention to a newer set, FrontierMath Erdős, built from 68 open problems. On that set, a pre-release version of Astra solved 2 in the official run.

The gap between the two numbers is the useful signal. Closing a curated set of very hard problems with known answers is a different task from making headway on open questions that nobody has solved. The second benchmark is where the next year of progress, or the lack of it, will show up.