GPT-5.4 Pro Closes a 60-Year-Old Erdős Problem , Math Research Just Changed

7. GPT-5.4 Pro Closes a 60-Year-Old Erdős Problem , Math Research Just Changed

Earlier this month, a combinatorics problem from Paul Erdős's open problem list , unsolved for 60 years , was solved with direct assistance from GPT-5.4 Pro. OpenAI researchers Sébastien Bubeck and Ernest Ryu discussed the result publicly on April 27, 2026, alongside host Andrew Mayne, framing it as evidence that frontier models have crossed a threshold in mathematical reasoning. No preprint has been linked yet, but the claim is specific: a named, decades-old open problem, a named model version.

This is the sharpest concrete data point in the ongoing race between OpenAI and Google DeepMind on formal and informal mathematical reasoning. DeepMind's AlphaProof and AlphaGeometry 2 demonstrated IMO-level performance in 2024, but those systems were purpose-built for formal proof search. GPT-5.4 Pro is a general-purpose model. If the claim holds under scrutiny, it suggests general reasoning capability has reached a point where no specialized architecture is required to contribute to research-grade mathematics. That shifts the competitive frame: the question is no longer whether a dedicated math AI can solve hard problems, but whether a general model already can. For Google, Anthropic, and xAI, this raises the bar on what a frontier general model is expected to do.

The next move worth watching is independent verification. Erdős problems vary widely in depth, and the specific problem has not been named publicly yet. Once identified, the mathematical community will assess whether the model's contribution was a key insight or scaffolding around human-directed proof steps. That distinction matters enormously for how this result gets priced into the competitive narrative around GPT-5.4 Pro and into hiring and research strategy at labs competing on reasoning benchmarks.

Source: @OpenAI on X