OpenAI and Anthropic Claim AI Breakthrough: What It Means

The story has moved beyond a pair of company announcements. On October 30, the news agency AFP published interviews with three prominent computer scientists connected to the hypothesis and the review effort. Their reactions—surprise, vindication and concern—show that the debate is now about both the mathematics and how AI-generated research should be checked.

What happened: OpenAI and Anthropic made similar claims

OpenAI announced its computer-science result on October 26, 2026, as part of a wider release covering mathematics, computer science and physics. The company said an internal AI system had produced an argument against the Exponential Time Hypothesis and that the work had subsequently been formalised in Lean, a proof-assistant system. OpenAI also said it wanted outside computer scientists to help examine and explain the result.

Anthropic followed with a separate paper and Lean 4 formalisation. Its repository says an internal Claude research model discovered the core algorithm. Anthropic’s paper says the result was found independently and concurrently with the OpenAI result.

• October 26, 2026OpenAI publishes its claim. The company releases the paper and formalisation and announces an academic collaboration to examine the computer-science result.
• October 27, 2026Anthropic publishes a parallel claim. Its research repository contains a paper and Lean 4 formalisation for an algorithm that it says refutes ETH.
• October 29, 2026Researchers describe the potential disruption. WIRED reports that academics are trying to understand what the claims would mean for decades of work built using ETH.
• October 30, 2026AFP publishes researchers’ first reactions. The original hypothesis proposers and MIT computer scientist Ryan Williams discuss the implications, uncertainty and access to AI research systems.

Read the original announcements: OpenAI’s research announcement, OpenAI’s ETH paper and Anthropic’s ETH research repository.

What is the Exponential Time Hypothesis?

The Exponential Time Hypothesis, usually shortened to ETH, is a conjecture about how difficult certain computer problems are. It is most commonly discussed through a problem called 3-SAT.

In 3-SAT, a computer is given a Boolean formula made from variables that can be true or false. The formula is arranged in clauses, with each clause containing up to three literals. The task is to determine whether there is any assignment of true and false values that makes the entire formula true.

A simple way to understand the challenge is to imagine trying every possible combination of values. With n variables, there are 2ⁿ possible assignments. More advanced algorithms can avoid checking every combination, but ETH says that no algorithm can solve every 3-SAT instance in genuinely subexponential time as the number of variables grows.

ETH has mattered because researchers have used it to establish conditional lower bounds: statements that certain algorithms cannot be much faster, assuming ETH is true. OpenAI says the hypothesis has supported hundreds of papers across areas such as graph algorithms, computational biology, databases and computational geometry.

What do the two AI companies say they found?

Both companies say their models found a deterministic algorithm with subexponential running time for a problem at least as demanding as deciding whether a 3-SAT formula is satisfiable. If the mathematical arguments hold as stated, that would contradict ETH.

The approaches are not identical, even though the headline conclusion is similar. The important point for readers is that these are mathematical claims about how an algorithm scales as an input grows—not demonstrations of a new consumer app or a faster setting that people can switch on today.

What are computer scientists saying on October 30?

The October 30 AFP interviews offer a more human picture of the story. Russell Impagliazzo and Ramamohan Paturi of the University of California, San Diego, helped formulate ETH with Francis Zane around 1999. MIT researcher Ryan Williams, who is also helping organise scrutiny of the claims, has publicly questioned the hypothesis in the past.

The researchers also raised questions about access. Powerful AI research systems are controlled by private companies, so publishing a proof does not automatically give the wider academic community equal access to the tools that generated it. Paturi said a company announcement cannot replace the research community’s own assessment.

At the same time, the interviews were not uniformly pessimistic. Williams described the period as a major change for theoretical computer science and encouraged students to learn how to work with AI tools while developing their own judgement.

Source: AFP’s October 30 interviews with computer scientists, republished by Tech Xplore.

Have the AI proofs been independently verified?

Not to the standard needed to call the result settled science. Both companies have made research materials public, and the work includes formalisation in Lean. That is meaningful: a proof assistant can check that a formal proof follows from the definitions and axioms encoded in the system.

But a machine-checked formal proof does not, by itself, answer every question a mathematician or computer scientist would ask. Reviewers still need to establish that the formal statements accurately represent the claims in the papers, that the modelling choices are appropriate, and that the reasoning has the intended mathematical meaning.

Anthropic’s own repository explains this distinction. It says Lean checks the formal proofs, while the correspondence between the short formalised statements and the paper’s statements still needs human inspection. The repository also describes an additional comparison check and a second proof-checking kernel. These are useful verification tools, but they are not the same as independent academic acceptance of the research result.

OpenAI says the computer-science work will be examined by a group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams. Their role is to help organise scrutiny and understanding; participation should not be mistaken for an endorsement. WIRED reported on October 29 that researchers had not yet independently verified the proofs or fully worked through their implications.

Does this mean AI solved P versus NP?

No. The Exponential Time Hypothesis and the P versus NP problem are related to computational complexity, but they are not the same question.

ETH concerns whether 3-SAT can be solved in subexponential time. P versus NP asks whether every problem whose proposed solution can be checked efficiently can also be solved efficiently. A subexponential algorithm can still take far longer than polynomial time. Therefore, even if the ETH claims are ultimately confirmed, they do not establish that P equals NP or solve the P-versus-NP problem.

Will this make computers faster or break encryption?

There is no immediate practical change for ordinary computer users. Anthropic’s own repository says the result is asymptotic and says nothing about solving formulas in practice. A theoretical algorithm can change the mathematical understanding of a problem without being useful on inputs that real systems encounter today.

The current claims also do not show that common encryption systems have been broken. The October 30 AFP interviews included a researcher saying the result itself is not a present security threat. Future developments in algorithms can always change long-term security assessments, but it would be misleading to tell users to change passwords, replace devices or abandon encryption because of this announcement.

What happens next?

The next important milestone is independent examination of the papers and formal proofs. Researchers will need time to understand the algorithms, compare the written papers with their formal statements, and identify exactly which existing results would need to be revisited if ETH is disproved.

Even if the claims hold, it does not follow that every paper relying on ETH becomes worthless. Some results may be conditional conclusions that remain valid under a different assumption; others may need new proofs or revised interpretations. The effect will depend on each result and on what the community confirms.

For now, the most accurate description is that two major AI labs have published parallel claims that could overturn a foundational assumption in theoretical computer science—and researchers are beginning the difficult work of checking what those claims really establish.

Frequently asked questions

Editorial note: This article reports claims made by OpenAI and Anthropic and reactions reported by AFP and other publications. It does not independently validate either mathematical proof. The verification status and any material developments should be updated when the research groups publish new findings.