Prime gaps at most 186, explained: what GPT-6 Astra proved and what it did not

On 30 August 2026, OpenAI published a short PDF titled Improved short gaps between primes. Its abstract states a single inequality, and one line beneath the theorem states something no mathematics paper has carried before: The proof is due to GPT 6 Astra. The result is a genuine record on how tightly consecutive primes can cluster, and it is real. It is also far narrower than the reaction suggested, and the distance between the claim and the coverage is the more interesting story.

What "prime gaps at most 186" actually says

Prime gaps at most 186 means this: there are infinitely many pairs of consecutive primes whose difference is no larger than 186. In the notation of the paper, the limit inferior of (p_ minus p_n) is at most 186, where p_n is the n-th prime. Put plainly, some gap size of 186 or less recurs infinitely often. The statement does not name which size recurs, it does not say where on the number line to look, and it says nothing about twin primes, the special case where the gap is exactly 2. The twin prime conjecture is the same sentence with 2 in place of 186, and after more than a century it remains unproven. A record on this bound is progress, not a solution.

Twelve quiet years, then three records in one month

The number has fallen like this:

When	Who	Repeated gap guaranteed at most
-	Twin prime conjecture	2 (unproved)
2013	Yitang Zhang	70,000,000
2013	Polymath 8a	4,680
2013	James Maynard, independently Terence Tao	600
2014	Polymath 8b	246
August 2026	Julia Stadlmann	240
August 2026	Axiom Math	212
30 August 2026	OpenAI, GPT-6 Astra	186

Zhang's 2013 work was the first unconditional finite bound, and it collapsed by five orders of magnitude within months as the Polymath group and Maynard's sieve picked at it. Then the number stopped moving. The 246 posted by Polymath 8b in mid-2014 stood for about twelve years. Julia Stadlmann, then at the University of Illinois Urbana-Champaign, broke the wall in August 2026 with arXiv 2608.31126 and a bound of 240, the first advance in over a decade. Three days later Axiom Math pushed it to 212. Within roughly two hours of Axiom's announcement, OpenAI reported 186.

A bigger sieve, not a bolt of lightning

The new proof is less a new idea than a rearrangement of known ones. It combines the equidistribution estimates of Polymath 8a and of Stadlmann with what the paper calls complementary factorization conditions, which make certain divisor products triply densely divisible. That extra divisibility buys a larger support for the multidimensional Selberg sieve, the machinery Maynard introduced in 2013, and with it a better numerical optimization. The payoff is concrete: an admissible set of 40 integers whose largest element is 186, a configuration the literature writes as DHL[40,2]. Because the set is admissible, every translate of it that sits far enough out must contain at least two primes, which forces infinitely many gaps of 186 or less. It is bookkeeping as much as inspiration, and that is precisely why a machine can now do it.

What the Lean file does and does not certify

OpenAI ran the argument through Lean 4, the proof assistant mathematicians use to check machine-written proofs, and published the repository as openai/PrimeGaps186, alongside a certificate script and a companion PDF of coefficient tables and verification records. The qualification matters more than the headline. The paper describes the formalization as conditional on numerical integral and cap bounds and on finite-field exponential-sum estimates that trace back to Deligne's work on the Riemann hypothesis over finite fields. Conditional means exactly that. This is not a proof verified end to end from the axioms; it is a long argument whose load-bearing numerical and analytic inputs are stated rather than discharged inside the formal system. The word is easy to lose in a headline, and it is the difference between a checked proof and a checked outline.

The gold goes to the human

The result landed in a community already arguing about what AI mathematics is for. Kannan Soundararajan of Stanford called the underlying problem "fiendishly difficult." Kevin Ford, Stadlmann's postdoctoral mentor at Illinois, described the summer as "a David-versus-Goliath story where the human gets the gold." Terence Tao was blunter, writing on mathstodon.xyz that what is happening is "a consequence of some very deliberate choices to abandon any pretense of gaining human understanding ... and using AI tools for the sole purpose of achieving a benchmark."

OpenAI's Sebastien Bubeck answered that charge in Science News' account of the month: "Our goal is not to preemptively strike and capture all the results possible," he said. "Our strategy is to empower the mathematician." Other mathematicians found the sequence of events harder to defend. Andrew Granville of Montreal said that "last week, I did a proof in two hours that would have taken me a month before," and of how the human record-holder was handled: "I do not see that OpenAI carefully considered how they treated her." OpenAI says it was in touch with James Maynard but did not contact Stadlmann directly. Her own response was pointedly unbothered by the number: "I am not so much interested in the exact number, but I am very interested in the mathematical ideas."

Why "OpenAI prime gaps" now means two papers

A search for OpenAI prime gaps now surfaces two papers from the same lab in the same week, concerning opposite ends of the problem. This article is about short gaps. The other, which we covered when the long-gaps claim surfaced, is about long gaps: the unusually wide stretches that open between consecutive primes. OpenAI published that work as a companion PDF, "Improved long gaps between primes," under the same 30 August date, and it too carries the line about GPT 6 Astra. The two are not interchangeable. OpenAI had been using the long-gaps problem as a benchmark for Astra, and the short-gaps result was described as an add-on pursued because the company's own mathematicians were curious. A record on the largest gaps says nothing about the smallest recurring ones, and only the second paper's title names the number the internet is now repeating.

What would settle it

The proof is public, the certificate tables are public, and the Lean development is public, which is more transparency than most AI proof claims get. Three things would move 186 from plausible to settled. First, a full writeup that discharges the conditional inputs, the numerical integral and cap bounds and the finite-field exponential-sum estimates, instead of citing them. Second, independent mathematicians reproducing the numerical certificate, which the companion PDF exists to make possible. Third, an unconditional Lean build with no assumptions left outside the system. Until then the honest description is the one OpenAI's own paper uses: conditional.

The record itself is modest in the scheme of things. From 246 to 186 is a real gain, not a revolution, and the twin prime conjecture is untouched. The method is the point. As one reply in the Hacker News thread on the repository put it, the result is simply that an infinite list of prime pairs differ by 186 or less. The thread itself stayed unmoved: 51 points, 11 comments, several dismissing the code comments as machine-written filler. That skepticism is the right default. A conditional certificate that a machine wrote is a starting point for mathematicians, not a conclusion for readers.