GPT-5.4 Pro cracks FrontierMath open problem

This is more interesting than a benchmark bump because it is a human-confirmed solve on a real open problem.

• –The task is construction-heavy combinatorics, where iterative search and verification loops can do real work because the answer is machine-checkable.
• –Epoch says Kevin Barreto and Liam Price first elicited the solution with GPT-5.4 Pro, then Will Brian confirmed it, so the signal is credibility as much as capability.
• –The page also says Opus 4.6 (max), Gemini 3.1 Pro, and GPT-5.4 (xhigh) later solved the same problem, so the signal is more about a newly reachable class of problems than a single model's monopoly on the result.
• –For AI developers, the practical takeaway is that agent-style workflows are starting to help discover candidate structures for real open problems, not just answer prompts.