OpenAI says its internal AI models have disproved the exponential time hypothesis, a computer science conjecture that researchers have leaned on for about 27 years. The company announced the result on Monday, October 26, 2026, and an Anthropic model has been credited with a parallel claim. Neither result has been peer reviewed, and independent experts are now checking both.

The hypothesis is a long-standing guess about how quickly computers can solve certain logic problems. It holds that some of those puzzles cannot be solved quickly by any computer program. According to the Washington Times, hundreds of papers depend on it, so a disproof would reach well beyond one lab. If the claim survives review, a working assumption that shapes how theorists think about the limits of computation would need rewriting.

What OpenAI and Anthropic say they found

OpenAI said on October 26 that its internal models had refuted the hypothesis, proved several mathematical conjectures and produced existence proofs for quantum field theories, according to Quartz. The company says its AI found a “natural language disproof” on Oct. 16. It formalized the result in Lean on Oct. 17, as Unite.ai reported.

The write-up is titled “The Exponential Time Hypothesis is false.” It describes a deterministic algorithm that exactly counts satisfying assignments of fixed-width CNF formulas in subexponential time and polynomial space, according to Technology.org. In plain terms, the algorithm counts how many ways a logic formula can be made true, and it does so in subexponential time. The formulas must have fixed width, which makes the scope of the result a central question for reviewers.

Anthropic's claim arrived hours after OpenAI's announcement. A paper credited to “Claude (Anthropic)” says an internal Claude research model found the core algorithm without human input on Oct. 16, the same discovery date OpenAI reports. FourWeekMBA dates the Anthropic paper to October 26. Two labs reporting the same discovery day is one of the open questions, and the sources have not reconciled the two accounts.

The claims also include a companion result. Forklog reports that an AI output dated October 18 proved the local smoothing conjecture, and that the proof was formalized in Lean on October 19.

OpenAI has an earlier episode on a far bigger prize. Nearly seven weeks before the October 26 announcement, the company floated an AI-generated proposed solution to the Navier-Stokes problem, according to R&D World. OpenAI now says it claims no scientific credit for this discovery, a stance that follows that earlier dispute, as BeInCrypto reports.

Why this is not a P versus NP result

The short answer to whether AI has solved P versus NP is no, based on what has been claimed so far. P versus NP asks whether every problem whose answer can be checked quickly can also be solved quickly. The exponential time hypothesis is a narrower question, and OpenAI's reported result concerns that narrower question. MindStudio reports that OpenAI's own president has said P probably still does not equal NP.

The distinction matters for the stakes. According to the Washington Times, if the hypothesis is false, some problems in fields such as graph theory, databases and computational biology may have faster solutions than experts assumed. OpenAI said finding them will take significant further research.

For algorithm designers, the near-term effect would be to reopen lines of work that the assumption had closed off. Graduate students who built thesis projects around hardness results would need to check which conclusions survive. Those are slow, specific consequences, not overnight ones.

Who is checking the claims

Russell Impagliazzo and Ramamohan Paturi will lead an independent review of the claim, with MIT's Ryan Williams, according to Technology.org. FourWeekMBA places Impagliazzo and Paturi at UC San Diego. The review is still pending, and no peer-reviewed paper has been published.

Alex Iosevich, a mathematics professor at the University of Rochester, called the result a “historic development,” as Quartz reported. Praise from outside experts is not the same as verification, and the review will decide which claims hold.

Lean checks whether a formal proof is valid. It cannot tell whether the formal statement says what the paper claims, and that judgment falls to human experts. OpenAI published the write-up and the formal proof on GitHub, and it has agreed to make all raw data public, according to the Washington Times.

In a statement quoted by the Washington Times, OpenAI said, “our system has recently proved that the [hypothesis] is false.” The company has not said which internal model produced the results, and that detail remains undisclosed.

The announcement also landed the same week as the GPT-6.1 Terra launch on October 26, per Technology.org, so the science story competed with a product launch for attention. OpenAI has taken a gated approach before. Its GPT-6 Rosalind biology model opened only to approved customers.

Does the claim threaten Bitcoin and encryption?

Not directly, according to the analysis so far. BeInCrypto reports that the result does not prove P equals NP, the larger question behind modern encryption, and that the algorithm poses no direct threat to the cryptography securing Bitcoin (BTC) and Ethereum.

The gap between a narrower complexity claim and a break in deployed cryptography is wide. Encryption and blockchain signatures depend on specific assumptions, and a claim that is still under review does not change them on its own. Readers searching for whether AI math proofs affect bitcoin encryption can take that answer for now.

No consumer action is required. Readers who want to follow the story can take three steps:

  • Wait for the independent review led by Impagliazzo, Paturi and Williams before treating the claim as settled.
  • Check whether the Lean formal statement matches the natural-language claim in the paper.
  • Watch for a peer-reviewed publication, which does not yet exist.

What to watch next

The next signals are the reviewers' verdict, whether OpenAI names the internal model behind its result, and whether the two October 16 discovery dates reconcile. Investor attention is also split, as USA 24 has tracked in its coverage of Big Tech's $3 trillion AI spending pledges.

If the review confirms both the algorithm and the Lean statement, expect a wave of papers testing which downstream results depended on the old assumption. If reviewers find a gap between the formal statement and the paper, the claim shrinks to something narrower. Either outcome will show how far AI-generated proofs can be trusted before humans have checked them.