Epoch AI reports GPT-5.4 Pro solved one FrontierMath Open Problems conjecture

Epoch says the solved item is a "Moderately Interesting" FrontierMath Open Problems entry: a conjecture from a 2019 paper by Will Brian and Paul Larson. In the follow-up post, Epoch says Brian had tried and failed to solve it both at the time of writing and in later attempts, which makes the claim more meaningful than a benchmark win on a synthetic task.

The linked problem page identifies the problem as a Ramsey-style hypergraph construction question and says GPT-5.4 Pro produced a solution that Brian confirmed. Epoch's publication update says Brian now plans to write up the result for publication, "possibly including follow-on work" generated by the model's ideas.

A context post quoting Brian's statement says the AI approach "works out perfectly" and "eliminates an inefficiency" in the prior lower-bound construction, which is the clearest public hint so far about what changed technically in the proof idea Brian quote screenshot.

Epoch is not presenting this as a single lucky prompt. In the main replication claim, it says Kevin Barreto and Liam Price first elicited a solution from GPT-5.4 Pro and that a third user did so shortly after, suggesting the result can be recovered by multiple operators rather than one irreproducible conversation.

More importantly for engineers building evals, Epoch says it reproduced the solve in its own scaffold. Its scaffold result says Gemini 3.1 Pro, GPT-5.4 xhigh, and Opus 4.6 max are all able to solve the problem "at least some of the time," which points to a harnessable capability rather than a model-specific anecdote.

Epoch also says the problem-page post includes a full chat transcript for GPT-5.4 Pro's original solution plus other models' solutions from the harness. That makes this unusually inspectable for a frontier-model math claim, even if the underlying success probabilities, prompt sensitivity, and verification costs are still not quantified in the thread.

The practical takeaway is that frontier models look increasingly useful as research assistants on hard formal problems, but only under heavy elicitation and verification. In the Tao discussion, Terence Tao describes current models as a "trustworthy coworker" for applying standard techniques, while also saying their broader hit rate on hard math remains around 1-2%.

That framing matters here. Epoch's own benchmark overview says only one of the four "Moderately Interesting" problems is solved so far, with every higher-importance tier still at zero. A reaction screenshot makes that breakdown explicit as 1/4 solved in the lowest tier and 0/11 in the more ambitious tiers scoreboard image.

So this story is best read as a reproducible milestone for evaluation design and assisted theorem discovery. It is not yet evidence that a lab can hand frontier models an open-problem queue and expect reliable autonomous research throughput.