OpenAI and Anthropic Models Claim Refutation of Exponential Time Hypothesis in Theoretical Computer Science

The sudden emergence of simultaneous, machine-generated challenges to a 30-year foundation in computational complexity tests whether corporate frontier models can overturn established theoretical frameworks before standard academic peer review can evaluate them.

• OpenAI published a refutation of the Exponential Time Hypothesis (ETH): On 26 October 2026, the lab announced that an internal model discovered a deterministic sub-exponential algorithm for counting satisfying assignments in fixed-width CNF formulas, verified partly via the Lean proof assistant.
• Anthropic issued a parallel claim hours later: The competing research lab stated that an internal Claude model had independently discovered a refutation earlier that month on 16 October 2026.
• Academic working groups mobilized immediately: Original formulators of ETH, including Russell Impagliazzo and Ramamohan Paturi, formed an expert group to independently audit the proofs, which currently lack formal peer-reviewed confirmation.

• The 30-year baseline: Proposed nearly three decades ago, ETH established that certain fundamental logical puzzles, such as variants of 3-SAT, cannot be solved substantially faster than exponential time in the worst case.
• Independence from P versus NP: While ETH helped complexity theorists map the precise limits of algorithm design, refuting it does not resolve the broader, unresolved P versus NP problem.
• Formalized verification methods: The partial encoding of these steps into the Lean interactive theorem prover represents an attempt to bypass traditional verification delays, moving mathematical proof-checking directly into code.

The sudden confrontation between corporate research laboratories and theoretical mathematics exposes a structural shift in how foundational knowledge is produced and vetted. For three decades, theoretical computer science operated on the premise that the Exponential Time Hypothesis accurately demarcated computational hardness. It provided algorithmic designers and complexity theorists with a stable benchmark: the mathematical certainty that worst-case instances of specific NP-complete problems demand exponential time to resolve.

By presenting a deterministic sub-exponential algorithm to count satisfying assignments in fixed-width CNF formulas, frontier AI labs are not merely proposing an algorithmic shortcut; they are challenging 30 years of derived complexity literature. Yet the underlying tension lies in the mechanics of disclosure. Historically, theoretical computer science advanced through open, peer-vetted dissemination, where proofs circulated through institutional channels and seminars long before public announcements. The release of simultaneous claims by OpenAI and Anthropic introduces corporate competition into pure theory, prompting criticism from researchers regarding "mathematics done by press."

This shift alters the incentives surrounding scientific validation. Corporate AI laboratories operate on development cycles measured in weeks, prioritizing speed and frontier model capabilities to establish technological preeminence. In contrast, academic complexity theory relies on exhaustive, slow-moving validation by a small cohort of global specialists. When complex mathematical results are partially verified through automated tools like the Lean proof assistant, traditional manual scrutiny becomes both more urgent and more difficult.

The immediate task facing specialists such as Russell Impagliazzo, Ramamohan Paturi, Fedor Fomin, and Leslie Goldberg is not just evaluating whether the logic holds, but auditing how automated theorem provers handle the nuanced edge cases of complexity theory. If the sub-exponential algorithm withstands this audit, the entire taxonomy of algorithmic lower bounds will require re-indexing. If the proof fails under scrutiny, the episode will underscore the risks of allowing proprietary model outputs to outpace standard academic consensus.

• Complexity theorists and academic mathematicians: Must immediately direct research bandwidth away from ongoing curricula to form auditing groups and dissect machine-generated Lean code.
• Frontier AI safety and evaluation teams: Face direct pressure to prove their automated theorem-proving pipelines produce sound mathematical foundations rather than complex hallucinations.
• Algorithm design specialists: Will need to recalibrate long-term assumptions regarding computational hardness and lower bounds if sub-exponential runtimes are validated.
• Corporate R&D leadership: Will evaluate whether formal verification tools like Lean can be deployed to automate the vetting of complex internal algorithms and theoretical claims.

• The Impagliazzo-Paturi working group audit: The release of formal technical commentary from the hypothesis’s original authors will signal whether the sub-exponential claim survives initial counter-analysis.
• Public release of complete formal proofs: Watch whether OpenAI and Anthropic publish full, unrestricted Lean codebases to allow general academic reproduction beyond their private blog posts.
• Specialized conference convening: Academic panels and open symposiums planned by the theoretical computer science community will mark the formal transition from press releases to peer-reviewed status.