Papex

Open problem · Mathematics · since 1849 · Open

Twin prime conjecture

Are there infinitely many pairs of primes that differ by 2?

Summary

Bounded gaps between primes have been proven since 2013, but not a gap of exactly 2 infinitely often.

AI-assisted claims

Partial result claimed · 16 Nov 2026

A narrower admissible 234 360 000-tuple and H_5 <= 4 891 341 734

“We construct an admissible 234360000-tuple of diameter 4891341734, which gives H_5 <= 4891341734 with Kapila's theorem.”

By
David Niedbala Giraudin
AI involvement
Text declaration: Primarily AI-generated.
Model
Anthropic Claude Opus, as stated
Where
Hexagon
Source last read
17 Nov 2026

Partial result claimed · 16 Nov 2026

An admissible 232 465 000-tuple with few long pairs, and H_5 <= 4 817 211 684

“We construct an admissible 232 465 000-tuple with 4 672 314 941 pairs further apart than 4 817 211 684, which gives H_5 <= 4 817 211 684 with the certificates of that preprint, unchanged.”

By
David Niedbala Giraudin
AI involvement
Text declaration: Primarily AI-generated.
Model
Anthropic Claude Opus, as stated
Where
Hexagon
Source last read
19 Nov 2026

Note: Papex lists version 1 of this preprint (claim 57b91304) with the weaker bound H_5 <= 4 891 341 734.

Partial result claimed · 04 Nov 2026

A sharper Type I/II estimate for primes in arithmetic progressions to smooth moduli, and H_m << e^{3.8043 m}

“As a consequence, liminf (p_{n+m} - p_n) << e^{cm} for every c > 3.80428, and for every c > 3.80423 if a numerically computed value of the sieve deficit is used. This improves Stadlmann's exponent 3.8075.”

By
David Niedbala Giraudin
AI involvement
Text declaration: Primarily AI-generated.
Model
Anthropic Claude Opus, as stated
Where
Hexagon
Source last read
05 Nov 2026

Partial result claimed · Revised · 04 Sep 2026

A fixed-profile proof for prime gaps of at most 182

“Her August 2026 mixed-support preprint lowered the more-than-twelve-year-old bound H1 ≤ 246 to H1 ≤ 240; independent concurrent work by OpenAI, using her 2023 estimates, reached H1 ≤ 186, and the Axiom Math team further developed her framework to reach H1 ≤ 212. Building on these developments with additional Type II and Type III inputs and a new fixed profile, this paper proves H1 ≤ 182.”

By
Romyar Sharifi
AI involvement
The paper is credited to “GPT-6 Astra (OpenAI), prompted by Romyar Sharifi”.
Model
GPT-6 Astra (OpenAI), as stated
Where
Romyar Sharifi's UCLA web page
Source last read
14 Oct 2026

Checks

3 author updates

  • 11 Sep 2026 · Author revision · Romyar Sharifi: “The 11 September 2026 version expanded citations and exposition; the stated results and numerical data were unchanged in that revision.” Source
  • 18 Sep 2026 · Author revision · Romyar Sharifi: “This edition incorporates corrections to the Type III result and adds its formal verification as a derivation from explicitly stated general published-theorem hypotheses.” Source
  • 03 Oct 2026 · Author revision · Romyar Sharifi: “Repository-publication references updated on 3 October 2026.” Source

Partial result claimed · 03 Sep 2026

A new bound for small gaps between primes

“Stadlmann recently proved that infinitely many pairs of distinct primes are apart by at most 240, improving upon the longstanding bound 246 established by Polymath8b. Building on her work, we prove that lim inf_{n→∞} (p_{n+1} − p_n) ≤ 212.”

By
François Charton, Letong Hong, Kenny Lau, Ken Ono, Guillaume Remy, Ho Chung Siu, Ashvin A. Swaminathan, Jesse Thorner, Yunzhou Xie (Axiom Math)
AI involvement
AxiomProver, an AI system under development by Axiom Math, autonomously generated from natural-language specifications a Lean certificate of the deduction of Theorem 1.1.
Model
AxiomProver, as stated
Where
Axiom Math website (paper PDF)
Source last read
14 Oct 2026

Partial result claimed · 03 Sep 2026

Improved short gaps between primes

“We prove that lim inf_{n→∞} (p_{n+1} − p_n) ≤ 186, where p_n denotes the nth prime.”

By
OpenAI
AI involvement
The proof is due to GPT 6 Astra. … Astra helped establish a stronger bound of 186, showing that infinitely many pairs occur within this smaller distance.
Model
GPT 6 Astra, as stated
Where
OpenAI (paper PDF), with Lean repository openai/PrimeGaps186
Formal proof
Lean 4, published by the source
Repository
openai/PrimeGaps186 at 61340d0 · read 21 Nov 2026
Statement file
Challenge.lean
Source last read
14 Oct 2026

Note: The repository README says the Lean results remain conditional on three explicit input axioms (two exponential-sum estimates from the literature and numerical integral bounds).

Claims as their sources state them. Listing a claim is not a check by Papex; the status stays the source’s.

All AI-assisted claims

Why it is open

Twin primes are pairs of primes p and p + 2. The question appeared in print in 1849, in a more general form due to Alphonse de Polignac, and in 1912 Edmund Landau listed it among four problems about primes that he considered "unattackable at the present state of science" [1]. Landau's list also contains Legendre's conjecture; János Pintz notes that the twin prime question concerns the smallest gaps between consecutive primes and Legendre's the largest [1].

Write H₁ for the smallest gap between consecutive primes that occurs infinitely often; the twin prime conjecture says that H₁ = 2 [9]. By the prime number theorem, consecutive primes near p are on average about log p apart [1], so the conjecture asks for gaps far below the average to keep recurring. A resolution would be a proof that twin primes never run out, or that they stop. This page's status follows the Wikipedia article on Landau's problems, which says that as of 2026 all four problems are unresolved [15].

Sieve methods, the standard tools for such questions, cannot tell apart integers with an even and an odd number of prime factors; Pintz calls the fact that Chen Jingrun's theorem cannot decide whether p + 2 is a prime or a product of two primes "the most significant particular case" of this parity problem, formulated by Atle Selberg [1]. The Polymath8b project showed that sieve-theoretic arguments of the kind behind today's bounds cannot go below a gap of 6, even assuming a strong conjecture on primes in arithmetic progressions, so a proof of the conjecture must go beyond them [4].

What is known

Early results (1900–1973). In 1900 David Hilbert mentioned the twin prime question in his eighth problem, together with Goldbach's problem and the Riemann hypothesis [1] (see the Papex pages Goldbach's conjecture and Riemann hypothesis). In 1920 Viggo Brun's new sieve showed that n and n + 2 both have at most nine prime factors for infinitely many n [1]. Chen proved (1966, 1973) that there are infinitely many primes p for which p + 2 is a prime or a product of two primes [1]. A 2026 preprint by Jiamin Li and Jianya Liu, not peer reviewed, claims a sharper version in which the smaller of the two factors is at most the 0.75th power of the larger [8].

The expected count. G. H. Hardy and J. E. Littlewood conjectured in 1923 that the number of twin prime pairs up to x is asymptotically C x/(log x)², with the twin prime constant C ≈ 1.3203; Andrew Granville writes that computational evidence suggests this is "a pretty good guess" [1, 6, 8].

Small gaps (1926–2009). Assuming the generalised Riemann hypothesis (GRH), Hardy and Littlewood showed in 1926 that gaps of at most about two thirds of the average size recur infinitely often; in 1966 Enrico Bombieri and Harold Davenport made such results unconditional by using the Bombieri–Vinogradov theorem in place of GRH [1]. That theorem says that primes are evenly spread over arithmetic progressions on average over moduli up to about √x [6]. Daniel Goldston, Pintz and Cem Yıldırım proved that gaps smaller than any fixed fraction of the average recur infinitely often, and that the Elliott–Halberstam conjecture, which extends the Bombieri–Vinogradov theorem to larger moduli, would give gaps of at most 16 [2, 6].

Bounded gaps (2013–2014). Yitang Zhang proved that H₁ < 70,000,000, using a stronger form of the Bombieri–Vinogradov theorem for moduli without large prime factors [3]; the online Polymath8 project lowered this to 4680 [4]. James Maynard, and independently Terence Tao, then found a multidimensional sieve that needs only the Bombieri–Vinogradov theorem; Maynard proved H₁ ≤ 600, and H₁ ≤ 12 under the Elliott–Halberstam conjecture, and showed that for every m some bounded interval contains m + 1 primes infinitely often [5, 6, 9]. Polymath8b reached H₁ ≤ 246, and H₁ ≤ 6 under a generalised Elliott–Halberstam conjecture [4]. These bounds are widths of admissible sets of shifts, sets for which no prime is bound to divide one of the shifted numbers; 246 is the width of the narrowest admissible set of 50 numbers [6, 9].

After 246 (2023–2026). Julia Stadlmann proved (preprint 2023, published 2025) that primes are evenly spread, on average, in progressions to smooth moduli up to x^(1/2 + 1/40), improving Polymath's 1/2 + 7/300, and deduced that intervals of length O(exp(3.8075m)) contain m + 1 primes infinitely often [7]. On 31 August 2026 she posted a preprint proving H₁ ≤ 240 by combining the Bombieri–Vinogradov theorem with such estimates; 240 is the width of the narrowest admissible set of 49 numbers, so she calls it "the smallest possible improvement", and she writes that she believes the method could achieve bigger ones [9].

Recent claims. On 3 September 2026 OpenAI released "Improved short gaps between primes", dated 30 August and credited to OpenAI, with its announcement of the model GPT-6 Astra [10, 14]. The paper claims H₁ ≤ 186 and states: "The proof is due to GPT 6 Astra" [10]. It describes Stadlmann's 240 as independent concurrent work, and its Lean 4 formalisation is conditional on numerical bounds and on exponential-sum estimates stated as assumptions [10].

A paper by François Charton, Letong Hong, Kenny Lau, Ken Ono, Guillaume Remy, Ho Chung Siu, Ashvin A. Swaminathan, Jesse Thorner and Yunzhou Xie, posted by Axiom Math as a preliminary draft dated 3 September 2026, claims H₁ ≤ 212 by sharpening Stadlmann's argument; it says that "The analytic ideas behind Theorem 1.1 originated from Stadlmann" [11]. Its appendix states that AxiomProver, "an AI system under development by Axiom Math", autonomously generated a Lean certificate of the deduction, which takes the analytic estimates and the variational certificate as hypotheses [11].

Romyar Sharifi posted on his UCLA web page "A fixed-profile proof for prime gaps of at most 182", dated 4 September and credited to "GPT-6 Astra (OpenAI), prompted by Romyar Sharifi", which claims H₁ ≤ 182, building on the three earlier bounds [12]. Revisions followed on 11 September (citations and exposition, results unchanged), 18 September (corrections to its Type III estimate, with a formal derivation from stated published-theorem hypotheses) and 3 October (references) [12]. The paper says Lean verifies its analytic-to-prime-gap step "under explicit numerical and finite-field premises" [12].

Tao wrote on 3 September of "the unedifying spectacle of no fewer than three separate AI companies race to announce their own improvement", noting that the results "are numerically stronger than Stadlmann's improvement" [13]. Science News quoted Stadlmann: "I am not so much interested in the exact number, but I am very interested in the mathematical ideas" [14]. On 14 October 2026 the Wikipedia article still listed the conjecture as unresolved and gave 246 as the bound [15].

Main approaches

Sieve weights for prime patterns. Take an admissible set of shifts and weight the integers n so that, on average, more than one of the shifted numbers is prime; then infinitely many translates contain two primes, at most the width of the set apart [5, 6]. Strongest evidence: bounded gaps, from H₁ ≤ 600 [5] to 246 [4] and 240 [9]. Main difficulty: the parity problem caps such arguments at 6 [4].

Primes in arithmetic progressions. A larger level of distribution, the range of moduli over which primes are known to be evenly spread, permits larger sieve supports and smaller bounds [6, 9]. Strongest evidence: Zhang's estimate for smooth moduli, which first gave bounded gaps [3], and Stadlmann's estimates [7], on which all the 2026 bounds draw [9, 10, 11, 12]. Main difficulty: the Elliott–Halberstam conjecture is unproven [6], and the newer estimates cover only restricted moduli; in Polymath's approach to H₁ that restriction cost more than it gained [9].

Large optimisations and machine certificates. The 2026 claims push the same sieve with larger numerical optimisations, written or checked with AI systems and Lean [10, 11, 12]. Strongest evidence: the claimed bounds 186, 212 and 182 [10, 11, 12]. Main difficulty: each Lean development takes analytic inputs or numerical certificates as hypotheses [10, 11, 12], and Tao argues that racing for such bounds favours answers over insight [13].

Zeros of L-functions. The first small-gap results assumed GRH [1, 6]. Roger Heath-Brown proved in 1983 that the existence of Siegel zeros, extreme irregularities in the distribution of primes in some arithmetic progressions, would imply the twin prime conjecture [1]. Strongest evidence: Pintz notes that both Siegel zeros and improvements of the Bombieri–Vinogradov theorem would imply bounded gaps [1]. Main difficulty: most mathematicians believe there are no Siegel zeros, and GRH would rule them out [1].

Open sub-questions

  1. Checking the 2026 bounds. Do the arguments for 186 [10], 212 [11] and 182 [12] hold as written, including the analytic inputs and numerical certificates that their Lean developments assume?
  2. Mixed supports. How far below 240 can Stadlmann's mixed-support method go with larger polynomial bases, which she expects to give bigger improvements [9]?
  3. Smooth moduli. Can the exponent of distribution for smooth moduli be raised above 1/2 + 1/40 [7], and what bound on H₁ would follow [9]?
  4. Many primes. Can the exponent 3.8075 in the bound for intervals containing m + 1 primes be lowered [7]?
  5. Past parity. What input outside sieve theory could prove a bound below 6, the limit of sieve methods even under the generalised Elliott–Halberstam conjecture [4, 1]?

Work on this in Papex

If you have a testable answer to one of the sub-questions above, you can draft it as a private hypothesis in Papex, or prepare the draft from Claude or ChatGPT with the Papex connector. You review the exact version, sign it and decide whether to publish it; a public record that names this problem is listed on this page. A Papex record shows who signed which version and when. It is not peer review and does not mean the hypothesis is correct.

Key references

  1. Pintz, J. (2009). Landau’s problems on primes. Journal de Théorie des Nombres de Bordeaux 21(2), 357–404. 10.5802/jtnb.676 · Checked 14 Oct 2026
  2. Goldston, D. A., Pintz, J. & Yıldırım, C. Y. (2009). Primes in tuples I. Annals of Mathematics 170(2), 819–862. 10.4007/annals.2009.170.819 · Checked 14 Oct 2026
  3. Zhang, Y. (2014). Bounded gaps between primes. Annals of Mathematics 179(3), 1121–1174 (received 17 April 2013). 10.4007/annals.2014.179.3.7 · Checked 14 Oct 2026
  4. Polymath, D. H. J. (2014). Variants of the Selberg sieve, and bounded intervals containing many primes. Research in the Mathematical Sciences 1, 12. 10.1186/s40687-014-0012-7 · Checked 14 Oct 2026
  5. Maynard, J. (2015). Small gaps between primes. Annals of Mathematics 181(1), 383–413. 10.4007/annals.2015.181.1.7 · Checked 14 Oct 2026
  6. Granville, A. (2015). Primes in intervals of bounded length. Bulletin of the American Mathematical Society 52(2), 171–222. 10.1090/S0273-0979-2015-01480-1 · Checked 14 Oct 2026
  7. Stadlmann, J. (2025). On primes in arithmetic progressions and bounded gaps between many primes. Advances in Mathematics 468, 110190 (first posted as arXiv:2309.00425 in 2023). 10.1016/j.aim.2025.110190 · Checked 14 Oct 2026
  8. Li, J. & Liu, J. (2026). Theorem (1 + 1.9) on the Goldbach Conjecture. arXiv:2606.05224 (v1 1 June 2026, v2 8 August 2026). Preprint, not peer reviewed. Source · Checked 14 Oct 2026
  9. Stadlmann, J. (2026). Bounded gaps between primes. arXiv:2608.31126 (v1, 31 August 2026). Preprint, not peer reviewed. Source · Checked 14 Oct 2026
  10. OpenAI (2026). Improved short gaps between primes. Manuscript dated 30 August 2026, released on 3 September 2026, with the conditional Lean formalisation in the GitHub repository openai/PrimeGaps186. Source · Checked 14 Oct 2026
  11. Charton, F., Hong, L., Lau, K., Ono, K., Remy, G., Siu, H. C., Swaminathan, A. A., Thorner, J. & Xie, Y. (2026). A new bound for small gaps between primes. Axiom Math, preliminary draft dated 3 September 2026. Source · Checked 14 Oct 2026
  12. Sharifi, R. (2026). A fixed-profile proof for prime gaps of at most 182. Expository bundle on his UCLA web page, edition of 18 September 2026 (original version 4 September 2026, references updated 3 October 2026); the paper is credited to "GPT-6 Astra (OpenAI), prompted by Romyar Sharifi". Source · Checked 14 Oct 2026
  13. Tao, T. (2026). Thread on Mathstodon, 3 September 2026 (posts 1–3), with the follow-up thread of 5 September 2026. Source · Checked 14 Oct 2026
  14. Mackenzie, D. (2026). A human set a new mathematical record. Then AI came for it. Science News, 11 September 2026. Source · Checked 14 Oct 2026
  15. Wikipedia. Landau’s problems (revision of 3 June 2026). Source · Checked 14 Oct 2026

Page reviewed 14 Oct 2026

Source

Open, as listed by Wikipedia · Landau’s problems. Checked 29 Sep 2026.

The status is the source’s. Papex does not decide whether a problem is solved.

Records on this problem

No public record addresses this problem yet.

More in Mathematics