OpenAI and Anthropic Challenge the Exponential Time Hypothesis With AI Research

The race to develop increasingly capable artificial intelligence systems is entering a new scientific phase, with OpenAI and Anthropic presenting research that challenges a long-standing assumption about the limits of computation. Their findings suggest that certain complex mathematical problems may be solvable more efficiently than computer scientists have traditionally believed.

The development, reported on October 27, 2026, has prompted renewed debate within the research community over the ability of AI systems to generate original scientific results and the procedures needed to verify them.

At the center of the discussion is the Exponential Time Hypothesis (ETH), a widely used assumption in theoretical computer science. OpenAI recently announced an algorithm that it says disproves this hypothesis, while Anthropic subsequently published its own work challenging the same principle.

Although the claims could have far-reaching implications, independent examination remains essential before the scientific community can determine their full significance.

Understanding the Exponential Time Hypothesis

The ETH helps researchers estimate how difficult certain computational problems become as their size increases. Under this hypothesis, some problems involving complex logical relationships cannot be solved efficiently because the time required grows exponentially as the number of variables increases.

This question is closely connected to the famous P versus NP problem, which asks whether every problem whose proposed solution can be verified quickly can also be solved quickly.

A familiar example is Sudoku. Checking whether a completed puzzle satisfies its rules is relatively straightforward, whereas finding a valid solution can require considerably more work. The distinction illustrates why verification and problem-solving are treated differently in computer science.

OpenAI's research reportedly introduces a deterministic algorithm capable of counting satisfying assignments for certain Boolean formulas in sub-exponential time. The result concerning the three-variable satisfiability problem, known as 3-SAT, is particularly significant because it would contradict the ETH if the proof withstands scrutiny.

However, disproving ETH would not automatically resolve the broader P versus NP question. A faster algorithm does not necessarily establish that every problem whose answer can be checked efficiently can also be solved in polynomial time.

Anthropic enters the debate

The announcement from OpenAI was followed by a separate contribution from Anthropic, intensifying interest in the possibility that advanced AI systems are becoming useful tools for research at the frontiers of mathematics and computer science.

The developments have also revived questions about how AI-generated discoveries should be assessed, credited and shared with the scientific community.

MIT computer scientist Ryan Williams is helping coordinate the examination of OpenAI's findings alongside researchers Russell Impagliazzo and Ramamohan Paturi of the University of California, San Diego. Impagliazzo and Paturi were among the researchers who introduced the ETH in 1999.

The group has called for wider scientific participation, including public discussion of the proofs and the standards that should govern AI contributions to academic research.

Williams has identified elements of the proposed solution that warrant particularly close examination. His participation is intended to help the research community understand the technical claims, rather than serve as a blanket endorsement of the results.

The process reflects a growing effort to combine the speed of AI-assisted discovery with independent mathematical verification.

Potential implications for cybersecurity

If the proposed algorithms prove correct, researchers may need to reconsider assumptions that underpin a substantial body of theoretical computer science.

The ETH has been used in numerous studies to establish limits on the efficiency of algorithms. A valid refutation could therefore require scientists to revisit earlier conclusions and explore new approaches to problems previously considered computationally difficult.

Cybersecurity is another area attracting attention. Many security systems depend on the practical difficulty of certain mathematical computations. Significant improvements in algorithms can sometimes weaken those assumptions, potentially affecting the security of encryption methods.

However, the findings do not currently establish that existing encryption systems have been compromised. Williams said OpenAI had examined whether its result created immediate cryptographic vulnerabilities and expressed confidence that the work did not imply such breaks.

Any practical consequences would depend on the precise capabilities of the algorithms, the problems they address and the security assumptions involved.

A new model for scientific research

Beyond the mathematical results themselves, the announcements highlight a broader transformation in the relationship between AI companies and academic researchers.

OpenAI has taken a more collaborative approach to reviewing some of its recent scientific claims, with external specialists helping organize the evaluation of work in computer science and physics. The objective is to ensure that findings with potentially wide-ranging consequences receive careful technical examination.

This approach also raises questions about transparency, access to research materials and recognition for the scientists involved in validating AI-generated discoveries.

For researchers, increasingly capable AI tools could accelerate work that once demanded years of specialized effort. At the same time, the ability to generate sophisticated proofs does not eliminate the need to establish that those proofs are correct, meaningful and genuinely advance existing knowledge.

The challenge may ultimately extend beyond verifying individual discoveries. If AI systems continue to tackle problems that have resisted conventional approaches, scientists could be compelled to rethink which questions deserve the greatest attention and how future researchers should be trained.

For now, the claims surrounding the Exponential Time Hypothesis represent a potentially important development rather than a settled scientific conclusion. Their lasting impact will depend on independent verification and on whether the proposed algorithms lead to further breakthroughs.