OpenAI GPT-5.4 Pro reportedly solves Erdős problem #1196 in 80 minutes, reveals novel mathematical connection

OpenAI's GPT-5.4 Pro model has reportedly solved Erdős open problem #1196 in approximately 80 minutes, according to discussions in the Erdős Problems forum. The model spent an additional 30 minutes formatting the solution as a LaTeX paper. Formal verification of the solution is currently underway.

Mathematical significance

Mathematician Terence Tao commented in the forum that the solution reveals a previously undescribed connection between the anatomy of integers and Markov process theory. "That would be a meaningful contribution to the anatomy of integers that goes well beyond the solution of this particular Erdos problem," Tao wrote.

Kevin Barreto, who says he will soon join OpenAI's AI for Science team, noted that the Markov chain technique used by GPT-5.4 Pro represented a creative step that human mathematicians had overlooked despite years of work on the problem.

Technical details

No specific details about GPT-5.4 Pro's architecture, parameter count, or training have been disclosed. The model's name suggests it is a variant within the GPT-5 series, though OpenAI has not officially announced this model publicly.

The solution is currently undergoing formal verification—a standard process in mathematical proofs where independent mathematicians check the validity of the work.

Implications for AI capabilities

The case provides evidence for ongoing debates about whether large language models can discover genuinely new knowledge beyond recombining training data. According to the source, the solution demonstrates that "new, previously undescribed knowledge can also be hidden within already known data points."

The Markov chain approach, while theoretically accessible to human mathematicians working with existing mathematical knowledge, was not identified by researchers who had worked on the problem for years.

What this means

If verified, this represents a concrete example of an AI system producing novel mathematical insight—not merely solving problems with known solution methods, but identifying new theoretical connections. The 80-minute solve time suggests the model engaged in extended reasoning or search processes, though the exact computational approach remains undisclosed. The significance lies less in automating mathematical proof and more in the model's ability to identify non-obvious connections between disparate mathematical domains that human experts had not recognized.