OpenAI says its AI disproved one of computer science's big assumptions

OpenAI says its internal AI models have refuted the Exponential Time Hypothesis (ETH), one of the most widely used assumptions in theoretical computer science. In the same announcement, the company claims its models proved a 35-year-old conjecture in mathematical analysis and produced a rigorous existence proof for quantum chromodynamics, work that also touches one of the Clay Mathematics Institute's Millennium Prize problems. None of the results has yet been reviewed by independent peers.

OpenAI published the announcement on Monday, October 26. The results will now be examined by external research groups, led in part by the researchers who originally proposed ETH.

What OpenAI claims

ETH is a conjecture about how fast computers can decide whether a logical formula can be made true, the so-called SAT problem. The problem is closely related to the famous P versus NP question. According to OpenAI, ETH has been used in hundreds of papers to show that tasks in graph theory, computational biology, databases and computational geometry cannot be solved substantially faster.

"Surprisingly, our system has recently proved that the ETH is false," OpenAI writes. The company says the model found the disproof on October 16 and that it was formally verified in the Lean proof assistant the next day. OpenAI has published both the model's natural-language proof and the Lean proof in its public GitHub repository.

OpenAI researcher Sébastien Bubeck writes on X that the result gives 3SAT in subexponential time, calling it "maybe the biggest surprise I have seen" in 20 years of following theoretical computer science.

Experts will check the work

The disproof will be reviewed by a group led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego and Ryan Williams of MIT. Impagliazzo and Paturi proposed the hypothesis together with Francis Zane. The group asked that all raw data be made public and that OpenAI fund a conference on the implications. OpenAI says it has agreed to both.

The mathematics result concerns Sogge's local smoothing conjecture from 1991. OpenAI says its models proved it in every dimension, which would also resolve the Kakeya conjecture and Stein's restriction conjecture. The physics result is an existence proof for theories including quantum chromodynamics and pure Yang–Mills theory. That work will be shepherded by Nobel laureate David Gross, and the group has decided not to release the AI-produced material until it has written up the work itself.

OpenAI does not seek academic credit for the discoveries, and members of the groups take part voluntarily, unpaid and without nondisclosure agreements.

What it means, and what it does not

OpenAI is clear that a theoretical breakthrough does not mean faster software tomorrow. Finding such algorithms and turning theory into practical improvements "will require significant further research," the company writes. The announcement says nothing about consequences for encryption or security.

For leaders, the main point is a different one: if the results hold, AI models have moved from solving known problems to overturning assumptions researchers have built on for decades. That raises questions OpenAI itself points to, about verification, access and who gets credit when a machine is behind a discovery.

Two caveats matter. The models that produced the results are internal and not available to customers. And the claims have not yet been independently confirmed, although Lean verification makes them far easier to check than ordinary proofs.

OpenAI ties the work to a $5 million contribution to the Simons Foundation, which will fund workshops and conferences where researchers can understand and build on results produced by AI.

Sources and media

• OpenAI: Working with academic communities to build scientific understanding
• Dan Roberts (OpenAI) on X: announcement of the three discoveries
• Sébastien Bubeck (OpenAI) on X about the 3SAT result

Source: OpenAI. The announcement is dated October 26, 2026. Dan Roberts' post on X was published on October 26 at 21:36 UTC. The sources were read by hogby.ai on October 27, 2026.

Illustration: AI-generated with Grok Bot / hogby.ai