No, Zhi-Wei Sun Did Not Set a Prime-Gap Record With GPT-5.6 Sol

A specific, named claim — a famous mathematician, a specific model, a world record — is the kind of detail that makes a story feel true. We went looking for the primary source and could not find it.

A version of this story has been circulating: Zhi-Wei Sun, the Nanjing University number theorist known for a long list of conjectures on primes and combinatorics, reportedly used OpenAI's GPT-5.6 Sol to set a new world record on prime gaps. It's a plausible-sounding story — Sun is real, prime gaps are a real and active area of number theory, and GPT-5.6 genuinely is credited with a real prime-gap improvement in August 2026. We checked the primary sources for the actual record, OpenAI's own materials, and Sun's own public output. None of them connect him to this result. Here's what we could verify, what we couldn't, and why the underlying pattern — a working mathematician using an AI model as a research tool — is still the more interesting story, even with the name wrong.

TL;DR — what we could and couldn't confirm

What prime gaps actually study

A prime gap is the distance between consecutive prime numbers. After 7 comes 11 — a gap of 4. After 89 comes 97 — a gap of 8. As numbers get larger, primes thin out and gaps grow on average, but never smoothly: some gaps stay tiny (twin primes, gapped by 2) and others stretch enormously, with no simple formula predicting which comes next. This irregularity is one of the oldest open frontiers in analytic number theory, and it splits into two genuinely different questions:

• How small can gaps get, infinitely often? This is the side that produced Yitang Zhang's celebrated 2013 breakthrough (bounded gaps under 70 million, later refined by the Polymath8 project — the same project Zhi-Wei Sun contributed an intermediate bound to) and remains open as the full twin prime conjecture: that a gap of exactly 2 recurs infinitely often.
• How large can gaps get? This is the side the August 2026 claim is actually about — proving that arbitrarily large gaps between consecutive primes are guaranteed to occur, and pinning down how large, as a function of the primes' size. The standing benchmark here since 2018 has been a bound proved by Ford, Green, Konyagin, Maynard, and Tao (FGKMT18), building on decades of work stretching back to Rankin in the 1930s.

Both questions are genuinely unsolved in their strongest forms. Nobody has proven the twin prime conjecture, and nobody has closed the large-gap question either — researchers only keep tightening the provable bound, the same way Claude tightened a Riemann zeta lower bound from 41.6% to 67.2% without touching the Riemann hypothesis itself.

The real record, as best we can source it

According to erdosproblems.com, the site mathematician Thomas Bloom maintains to track progress on Erdős's catalogued open problems, an entry updated around August 30–31, 2026 records an improvement to the FGKMT18 large-gap bound. The new argument reportedly removes a log log log n factor from the denominator of the previous formula, yielding gaps of size proportional to:

(log log n) / (log log log log n) · log n

for infinitely many consecutive primes — a genuine, if incremental by number-theory standards, tightening of an eight-year-old record. The page's own attribution line, per available reporting, credits the underlying idea to "GPT-5.6 Pro (prompted by DottedCalculator)" — a pseudonymous account, not a named professional mathematician — combined with the existing FGKMT18 sieve machinery, with Bloom himself writing the full expository proof. Coverage of the episode also describes the new construction as using "skewed residue classes," a variant sieve-theory technique distinct from the standard Maynard-weight approach.

The result is described in multiple write-ups as "purely elementary" — built from ideas "that could have been discovered decades earlier" — and as working only in combination with the existing FGKMT18 machinery, not as a standalone new proof.

What we found no evidence of: any mention of Zhi-Wei Sun in the erdosproblems.com attribution, in OpenAI's own PDF write-up of a related long-gaps result, or in independent coverage of the August 30 claim. Sun does have a genuine, decades-long connection to prime-gap research — beyond the 2013 Polymath8 contribution, he has separately posted open problems and bounties on erdosproblems.com on unrelated questions. That real history may be exactly why a version of this story attached his name to the 2026 AI result. We're flagging it rather than repeating it.

What's also unresolved, independent of attribution: how thoroughly this specific proof has been checked. Some coverage states the result was "rigorously verified through formal machine proof." Closer reporting disputes that framing directly, noting that the erdosproblems.com page records only that the statement of the result is formalized in Lean — not that the full multi-step argument passed a Lean kernel check the way Claude's Fermat's Last Theorem formalization or the Riemann zeta bound's Lean certificate did. That's a meaningful gap between "verified" and "the claim about verification is itself verified."

The story that is well documented: a doctor, not a lab

If the headline version of this story is shaky, a close cousin of it checks out cleanly, and it's arguably the more important pattern for anyone thinking about AI adoption.

In August 2026, Jin Shanmu, a neurosurgery resident at Peking Union Medical College Hospital in Beijing with an undergraduate degree in geology and no formal advanced-mathematics training, used GPT-5.6-Sol running unsupervised for roughly 16 hours on ChatGPT to produce a complete manuscript proof of Crouzeix's conjecture — a 2004 open problem in matrix analysis stating that applying any function to a matrix never inflates its norm by more than a factor of two relative to the function's maximum on the matrix's numerical range. Jin reportedly encountered the conjecture while researching an unrelated topic, transcranial ultrasound. Michel Crouzeix, the conjecture's original author, reviewed and confirmed the proof's correctness — with the caveat, worth keeping honest, that one review by the problem's own author is not the same thing as broad, independent peer review.

That story matters for a different reason than a lab's benchmark run. It's the same distinction we drew when covering Claude's non-mathematician operator on the Riemann result, where an Anthropic staffer's contribution was largely "keep going" and "believe in yourself" — except here the human isn't even inside the lab that built the model. Jin is exactly the kind of user the "AI helps ordinary researchers" thesis is supposed to be about: someone with no domain credential, using a public product tier, on a problem he stumbled into sideways. If that pattern generalizes, it says more about adoption than any single lab showcase does — because a lab choosing which win to publicize is a curated sample of one, while an unrelated professional independently stumbling into a real result with the same tool is a much harder thing to manufacture.

Why "a mathematician used it as a tool" would matter more than "a lab solved it"

This is the actual reason the Zhi-Wei Sun framing was worth chasing down, even after it didn't check out. Every AI-math result explainx.ai has covered this year up to now shares one structural feature: a lab is the actor. Anthropic ran the Riemann zeta search and formalized Fermat's Last Theorem with its own unreleased research models. OpenAI's own team produced the Erdős planar unit-distance result. Fable 5 produced the Jacobian conjecture counterexample with a named Anthropic-adjacent mathematician, Levent Alpöge, validating it. Even the "660 subagents and a cheerleader" texture of these stories is a lab controlling the entire pipeline — model, compute, timing, and the announcement itself.

A working mathematician (or, in Jin's verified case, a total outsider) opening a public product tier and getting a genuine result is a different and more consequential claim, because it's not selectable by a PR team. Labs publish their wins; they don't publish every failed run. An individual researcher using the same public tier that thousands of other people are also using, on a problem they picked for their own reasons, is closer to what real adoption looks like — mundane, distributed, and much harder to cherry-pick. That's precisely why we wanted the Zhi-Wei Sun version of the prime-gap story to be real: a globally respected number theorist folding GPT-5.6 Sol into his actual research workflow would be a stronger adoption signal than another lab showcase. It just isn't the story that's actually documented — Jin Shanmu's Crouzeix's conjecture case is.

What this means if you're evaluating AI-math claims yourself

1. Check who the primary source actually credits, not just who a headline names. erdosproblems.com's own attribution line is the ground truth for the prime-gap record, and it names a pseudonymous prompter, not a public figure.
2. "Formalized in Lean" is not one fact. Ask whether the statement was formalized (weak) or the full proof passed a kernel check (strong) — the gap between those two claims is exactly what got blurred in secondary coverage of this record, and exactly what the Fermat's Last Theorem and Riemann zeta posts on this site spell out for their own results.
3. A real, respected name attached to an unrelated real event is a classic misattribution pattern. Sun's genuine 2013 Polymath8 contribution to bounded gaps is likely why his name reads as plausible next to a 2026 large-gaps claim — the two are real but separate events, decades and directions apart.
4. The more mundane story is usually the more trustworthy one. A neurosurgery resident quietly proving a 22-year-old conjecture while researching something else entirely required no PR timing to be checkable — Crouzeix himself confirmed it.

Bottom line

We could not verify that Zhi-Wei Sun set a prime-gap world record using GPT-5.6 Sol, despite Sun being a real, prominent number theorist with a genuine, separate history in prime-gap research. What is documented: an August 30, 2026 improvement to the 2018 FGKMT large-gap bound, credited by its own primary source to "GPT-5.6 Pro" prompted by a pseudonymous account, with mathematician Thomas Bloom writing the exposition — and a formalization claim that's narrower than some coverage suggests. The pattern worth actually tracking — an ordinary researcher using a public AI model as a daily tool, not a lab running its own showcase — is real this year. It just belongs to a Beijing neurosurgery resident and a 2004 conjecture about matrix norms, not to Zhi-Wei Sun and prime gaps.

Related on explainx.ai

• Claude formalized Fermat's Last Theorem in Lean 4 — 13M lines, 29,500 theorems — today's other AI-math result, and the sharper reference for what "machine-verified" should mean
• Claude pushed a Riemann zeta lower bound from 41.6% to 67.2% — the framing this "AI extends existing human results" pattern follows
• OpenAI's models resolved an 80-year Erdős planar unit-distance problem
• GPT-6 Astra launch: every number that actually matters — the model family this record's successor tier belongs to
• Did Fable 5 disprove the Jacobian conjecture? The Alpöge thread explained
• Star Fleet Math: 20 parallel Codex agents chase 27 Erdős proposals in Lean 4
• Will AI replace mathematicians? The IEEE "Big Mathematics" debate

Primary sources checked: erdosproblems.com · OpenAI's PDF on improved long gaps between primes · Zhi-Wei Sun, Wikipedia · reporting from Traictory and AI Weekly on the prime-gap claim and the separate Crouzeix's conjecture case

Accurate as of September 5, 2026. We could not independently confirm the Zhi-Wei Sun attribution despite checking erdosproblems.com, OpenAI's own materials, and Sun's public academic profile; if a primary source surfaces connecting him to this specific 2026 result, we will update this post. Follow @explainx_ai for updates.