OpenAI and Anthropic Models Claim Refutation of Exponential Time Hypothesis Under Expert Review

The collapse of a foundational theoretical computer science pillar would invalidate hundreds of academic lower-bound proofs across graph theory, computational biology, and database design while shifting the frontier of scientific discovery into proprietary AI labs.

• The formal challenge: On 26 October 2026, OpenAI published claims that internal AI models generated a deterministic sub-exponential time algorithm to count satisfying assignments in fixed-width k-CNF formulas, refuting the nearly 30-year-old Exponential Time Hypothesis (ETH).
• Parallel discovery: Hours after OpenAI's release, Anthropic published an independent refutation derived from its Claude models, demonstrating simultaneous automated discovery across competing frontier architectures.
• Academic mobilization: The claims, formalized partly using the Lean interactive proof assistant, are undergoing review by an independent expert group that includes original ETH co-creators Russell Impagliazzo and Ramamohan Paturi alongside MIT's Ryan Williams.

• The mid-1990s foundation: Formulated roughly three decades ago, ETH conjectures that NP-complete problems like 3-SAT require exponential time to solve, functioning as a practical baseline for hardness without proving or resolving the broader P versus NP problem.
• The Lean verification pipeline: The discovery was captured by models on 16 October 2026 and encoded into Lean code, shifting scientific validation from narrative mathematical prose to interactive proof checking.
• The hundreds of exposed theorems: Academic literature relies on ETH as a load-bearing assumption; hundreds of published papers use it to prove that specific algorithms in geometry, networks, and biology cannot run any faster than current theoretical thresholds.

The technical debate centres on a clean operational mechanism: counting satisfying assignments for k-CNF formulas (known as #k-SAT). Under the classical formulation of the Exponential Time Hypothesis, computer scientists accepted the premise that algorithms solving this family of problems cannot run in sub-exponential time relative to the number of variables. This assumption became an analytical shortcut. Rather than proving lower complexity bounds from first principles—an unresolved challenge in complexity theory—researchers indexed their proofs to ETH. If ETH held, thousands of downstream problems across network routing, genomic alignment, and geometric packing were mathematically guaranteed to have fixed computational speed limits.

The announcements from OpenAI and Anthropic attack that baseline directly. By producing a deterministic sub-exponential algorithm, the models assert that this fundamental complexity ceiling was drawn in the wrong place. The machine-generated proofs utilize the Lean proof assistant, an interactive theorem prover designed to confirm the logical validity of mathematical steps. While human review remains necessary to verify whether the formalization accurately maps to the full theoretical statement of ETH without subtle scope limitations, the integration of automated theorem provers alters the verification cycle.

This shift creates immediate friction in the incentives governing academic research and frontier artificial intelligence labs. For theoretical computer scientists, foundational results have traditionally emerged through transparent, multi-year peer review. In this instance, two private corporate labs presented simultaneous refutations developed inside proprietary models, prompting critiques regarding "mathematics done by press." However, the dual discovery also reflects a structural change: frontier foundation models are transitioning from code assistance and natural language synthesis to traversing formal symbolic search spaces that human intuition abandoned decades ago.

Crucially, the practical boundaries of computational infrastructure remain stable in the immediate term. Cryptographic frameworks—such as public-key infrastructure and lattice-based post-quantum protocols—do not rely on ETH for security assurances, as they depend on average-case hardness rather than worst-case NP-completeness. Instead, the downstream implications concentrate inside theoretical algorithm design. If the sub-exponential algorithm withstands expert scrutiny, computer scientists must re-evaluate hundreds of published hardness results, forcing a comprehensive audit of modern computational complexity textbooks.

• Theoretical computer scientists: Researchers at institutions like UC San Diego, MIT, the University of Bergen, and Oxford face immediate line-by-line verification workloads, with existing conditional hardness proofs facing obsolescence if the refutation stands.
• Frontier AI enterprise labs: Engineering and research leads at OpenAI and Anthropic secure technical validation for automated reasoning architectures, shifting research-and-development investment toward formalized theorem proving in Lean.
• Algorithm architects in bioinformatics and logistics: Software engineers relying on theoretical worst-case ceilings for NP-hard optimization problems may eventually access new design patterns for fixed-parameter tractable algorithms.
• Graduate faculties and academic publishers: University departments and journal editors confront the logistical burden of auditing, retracting, or rewriting core curricula and literature that took ETH as an absolute premise.

• The working group formal report: Watch the joint evaluation from Impagliazzo, Paturi, and Williams, as their formal assessment will confirm whether the Lean proofs map precisely to the full mathematical formulation of ETH or address a narrower sub-problem.
• Lean code repository validation: Track community verification of the formal Lean scripts, which serves as the deterministic test of whether the AI models' step-by-step logic contains formal semantic errors or axiomatic hallucinations.
• Textbook and curriculum revisions: Monitor updates to foundational computer science texts, signaling that the theoretical computing community has officially retired ETH as a baseline conditional assumption.