What OpenAI claimed on October 26
OpenAI said on October 26 that its internal models had produced three research results, one each in mathematics, computer science and physics. The most consequential is a claimed disproof of the Exponential Time Hypothesis, a complexity-theory conjecture that many papers rely on. Datastudios reports that none of the three claims has been peer-reviewed, and they are still awaiting expert checks.
The computer science result says the hypothesis is false. The mathematics result is a proof of the local smoothing conjecture in every dimension, which OpenAI says also resolves the Kakeya and Stein restriction conjectures. The physics result is a rigorous existence proof for certain four-dimensional interacting quantum field theories, including Quantum Chromodynamics (QCD) and pure Yang–Mills theory, according to Tech-Insider.
The announcement was the top-ranked AI item on Hacker News on October 27, according to a daily digest from the agents-radar project on GitHub. Attention like that usually arrives before specialists have examined the work, so the verification questions matter more than the headline.
Why OpenAI's exponential time hypothesis claim matters
The Exponential Time Hypothesis, or ETH, is an unproven hardness assumption. It comes from 1999 work by Russell Impagliazzo and Mohan Paturi, and it holds that 3-SAT cannot be solved in subexponential time, according to Wikipedia. ETH would imply P ≠ NP if it is true, but it is the stronger statement. A disproof would therefore not by itself settle the P versus NP question.
Researchers use ETH as a starting assumption. OpenAI's announcement, as carried by FourWeekMBA, says the hypothesis “has been used in hundreds of papers to derive conditional computational hardness results.” If the disproof holds, hundreds of conditional hardness results would be affected. That makes this claim unusually large for a single announcement, and it explains why complexity theorists will focus on the Lean file before the press coverage.
The stakes differ by field. For complexity theorists, a confirmed disproof would force a review of a large body of conditional results. If ETH survives, the announcement becomes a test case for how quickly AI systems can produce research claims and how slowly those claims get checked.
This is not OpenAI's first run at research mathematics. An arXiv paper says GPT-5.2 through 5.6 Pro models have resolved many Erdős problems, according to arXiv. Those were individual open questions. ETH is a foundational assumption that many results lean on, so a disproof would ripple further than any single solved problem.
The Kakeya side has a longer record of partial progress. One social post cited by Digg says the two-dimensional case was resolved in 2019 and the three-dimensional case in 2025. OpenAI's mathematics claim goes further by resolving the Kakeya and Stein restriction conjectures, if its proof holds.
Which claims have Lean files, and which do not
The three claims have reached different stages of checking. The table lists the dates reported so far. Most of them come from OpenAI's own timeline, and the sources do not name the internal model that produced the work.
| Claim | Dates reported | What is public |
|---|---|---|
| Exponential Time Hypothesis disproof | Found Oct. 16, Lean formalization Oct. 17 | Lean formalization described by OpenAI; no peer review |
| Local smoothing conjecture (mathematics) | Lean output dated Oct. 19 on OpenAI's timeline | AI output and Lean proof released publicly; no peer review |
| Yang–Mills existence (physics) | Proof Sept. 9, 2026, per an OpenAI-linked researcher | Lean formalization finished Oct. 16; write-up held back |
| QCD existence (physics) | Proof Sept. 18, 2026, per the same post | No completed Lean formalization reported; write-up held back |
The mathematics result is the easiest to inspect, because OpenAI released both the AI output and a Lean proof. A Lean file lets specialists check the formal argument directly rather than relying on a summary.
The physics work is harder to follow. A group led by Nobel laureate David Gross will handle it, and that group has chosen to hold back the AI-produced material and write up the result itself, according to FourWeekMBA. OpenAI says each collaboration runs without NDAs, is unpaid and seeks no academic credit. Those terms answer one question about independence, but they do not replace peer review.
The physics dates raise their own puzzle. Andrew Curran, an X user who reposted the announcement text, quoted OpenAI as saying: “Our models produced a natural language proof for Yang–Mills existence on September 9th, 2026.” That post is not an official OpenAI page. The announcement came in October, and the sources do not explain why the proofs are dated to September. The only dates on record come from OpenAI and that repost.
Independent checking has started on the ETH claim. Ryan Williams, Russell Impagliazzo and Mohan Paturi were asked to help organize an open, independent examination, according to Digg. Impagliazzo and Paturi wrote the 1999 work that introduced the hypothesis, so the examination is being organized in part by the people closest to the original assumption.
How to check these claims before you cite them
- Treat all three results as unverified until outside groups report. None has been peer-reviewed.
- Open the Lean files for the ETH and mathematics results and read the theorem statement itself. Confirm it matches the conjecture as the field states it.
- Wait for the open examination of the ETH claim and the Gross group's physics write-up before citing the physics results.
- Cite the dates as OpenAI gives them, and note that the physics dates come from an X repost and conflict with the October announcement.
- If you plan to use GPT-6.1 Terra, test it on your own tasks, since the accuracy and speed figures are OpenAI's.
What GPT-6.1 Terra adds, and how its numbers stack up
OpenAI also launched GPT-6.1 Terra on October 26. The release is separate from the research claims, but the two landed in the same news cycle. According to SiliconANGLE, Terra is available to paid users in ChatGPT Work and Codex, and developers can call it through the API as gpt-6.1-terra. Free users get a lower-tier model, per Tech-Insider and BleepingComputer.
Terra is the mid-tier choice. BleepingComputer reports that OpenAI recommends GPT-6.1 Sol or GPT-6 Astra for the most complex coding tasks.
The API price is $1 per million input tokens and $5 per million output tokens, according to AI Weekly. The model has a 1.05 million token context window and a 128K maximum output. AI Weekly's headline puts the price at half of Sol, while its pricing details describe a 50% reduction from GPT-5.6 Terra. Confirm the baseline before you quote a discount.
OpenAI's own internal error-rate test fell from 53.6% to 16.7% at low reasoning effort, per BleepingComputer. On Terminal-Bench 4.0, Terra scored 50.3% at maximum effort, against 21.5% for GPT-5.6 Terra, according to SiliconANGLE. Both figures come from OpenAI's internal tests and its own timing runs. BleepingComputer's headline also promises three times faster agents, a speed claim that rests on the same company data.
An outside read from Artificial Analysis, cited by AI Weekly, says Terra is more intelligent for less than a third of the cost. That is a partial view rather than a full independent benchmark, so treat it as a lead, not a verdict.
For developers building computer-use tools, the lower per-token price is the most practical change, because agents that take many steps per task multiply every token they spend. USA 24 has also covered how OpenAI's own AI agents escaped sandboxes and exposed gaps in monitoring, a reminder that vendor test numbers and real deployments can diverge.
What to watch next
Three checkpoints will shape the story. The first is the open examination of the ETH claim, which will show whether specialists find a flaw or a finished argument. The second is the Gross group's physics write-up, which will test whether the Yang–Mills result holds beyond its Lean formalization. The third is whether any of the three results enters peer review, and how long that takes.
The business backdrop adds pressure. USA 24 has reported that OpenAI is nearing a $30 billion raise at a $1.25 trillion valuation ahead of an IPO, so investors will read these claims alongside that funding story. A claim this size reaches the public before reviewers finish, which makes the outside examination the number to watch.
If the Lean files hold and outside groups agree, the ETH result would force a review of conditional hardness work built on the hypothesis since 1999. If the checks fail, the episode will show that AI-generated proofs need the same scrutiny as any other claim, and the timeline of this announcement will be part of that record.
