Skip to content
← All stories AI Research & Models

OpenAI models claim to refute the Exponential Time Hypothesis

OpenAI’s frontier labs released a proof that overturns the Exponential Time Hypothesis, a long‑standing assumption in computational complexity, prompting swift reactions from Anthropic and MIT researchers.

3 min read

OpenAI’s frontier labs announced on 27 October 2026 that their AI models have produced a proof refuting the Exponential Time Hypothesis (ETH) (TechCrunch, “Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?”, 27 October 2026).

How did OpenAI claim to refute the Exponential Time Hypothesis?

OpenAI posted a pre‑print on GitHub that describes an algorithm solving ETH‑class problems in sub‑exponential time, which the lab says directly disproves the hypothesis (TechCrunch, “Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?”, 27 October 2026). The same pre‑print was linked in an OpenAI community post on 27 October 2026, confirming the release date (OpenAI Community, “The Exponential Time Hypothesis is False”, 27 October 2026). The proof was built and checked in roughly ten days, according to a summary on MindStudio (MindStudio, “AI Just Disproved the Exponential Time Hypothesis. Here’s Why It …”, n.d.).

What is the Exponential Time Hypothesis and why is it important?

The Exponential Time Hypothesis, introduced in 1999 by Impagliazzo and Paturi, asserts that certain constraint‑satisfaction problems cannot be solved in less than exponential time, providing a foundation for believing P ≠ NP and for the security of many cryptographic protocols (TechCrunch, “Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?”, 27 October 2026). ETH has been cited in countless papers over the past two decades, so a refutation would force a re‑evaluation of theoretical computer‑science assumptions and potentially affect the design of algorithms that rely on exponential‑time hardness.

How are other AI labs and researchers responding?

Anthropic released its own ETH refutation on GitHub within hours of OpenAI’s announcement, echoing the claim that their models found sub‑exponential solutions (TechCrunch, “Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?”, 27 October 2026). MIT professor Ryan Williams, who helped formulate the original ETH, organized a working group of computer scientists to scrutinize the proofs, describing OpenAI’s earlier “proof dump” as a “DDOS attack on mathematicians” but agreeing to collaborate on interpretation (TechCrunch, “Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?”, 27 October 2026). UC San Diego researchers Russell Impagliazzo and Ramamohan Paturi also joined the effort, underscoring the community’s willingness to engage with the AI‑generated results.

What are the potential implications for cryptography and computational theory?

If the sub‑exponential algorithms are correct, cryptographic schemes that rely on the assumed exponential hardness of certain problems could become vulnerable, though OpenAI’s team said they have not identified any immediate breaks in widely used protocols (TechCrunch, “Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?”, 27 October 2026). The discovery also fuels speculation that the boundary between P and NP may be narrower than previously thought, but experts caution that the result is unlikely to resolve the P vs NP question outright (TechCrunch, “Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?”, 27 October 2026).

Timeline of key events

DateEntityAction
25 October 2026OpenAIPosted pre‑print proof refuting ETH on GitHub
27 October 2026OpenAIAnnounced breakthrough via frontier labs (TechCrunch)
27 October 2026AnthropicReleased its own ETH refutation on GitHub
27 October 2026MIT (Ryan Williams)Formed working group to assess proofs

Sources

openai exponential-time-hypothesis computational-complexity anthropic mit