OpenAI caught overhyping GPT-5's math skills

If you saw the viral posts claiming GPT-5 had just solved math problems that stumped humanity for decades, take a deep breath. The entire breakthrough unravelled in less than a day, turning into an awkward public relations mess for OpenAI.

It all kicked off when OpenAI manager Kevin Weil posted on X that GPT-5 had found solutions to ten previously unsolved Erdős problems and made progress on eleven more. Other OpenAI researchers amplified the claim, using language that made it sound like the upcoming frontier model had independently cracked open tough number theory questions.

A dramatic misinterpretation

Mathematicians were quick to check the receipts. Thomas Bloom, who runs the erdosproblems.com database cited in the claims, pushed back publicly against the announcement.

Bloom called the statements a dramatic misinterpretation. He pointed out that an open tag on his website simply meant he personally did not know the solution or had not logged it yet, not that the mathematical community had never solved it. GPT-5 had not invented novel proofs. It had simply dug up obscure academic papers that Bloom had missed.

Once the mistake became obvious, the tweets were mostly scrubbed, and the researchers walked back their claims. But the damage was done. DeepMind CEO Demis Hassabis called the communication embarrassing, while Meta AI chief Yann LeCun mocked OpenAI for getting hoisted by its own hype.

What GPT-5 actually did right

The irony is that underneath the exaggerated marketing, GPT-5 actually did something genuinely impressive. It acted as an ultra-capable literature review engine, connecting scattered papers where terminology differed across decades.

Mathematician Terence Tao weighed in with a grounded perspective. Tao noted that AI's most immediate and practical role in mathematics is not cracking elite unsolved conjectures, but dramatically accelerating tedious tasks like cross-referencing academic archives. While generative AI could help industrialize mathematical workflows, Tao stressed that rigorous human expertise remains essential to verify and classify whatever the model surfaces.

Why it matters

We think this slip-up exposes just how much pressure OpenAI is facing to prove that frontier models are evolving into autonomous scientific engines. With billions on the line, even top managers fell into the trap of using loose phrasing to turn an effective literature retrieval into a fake scientific milestone.

For researchers and knowledge workers, the takeaway is straightforward. Do not expect GPT-5 to magically solve your hardest theoretical roadblocks right now. However, treating it as an exceptionally fast research assistant for unearthing buried papers is a genuine productivity leap worth adopting today.

The industry does not need another cycle of self-inflicted hype. Precision and transparency about what AI can actually do will earn far more trust than manufactured breakthroughs.