After 78 years and one famous false start, the six-dimensional sphere S⁶ finally has a complex structure. Harvard mathematician Levent Alpöge and Anthropic's Claude constructed it explicitly in a 108-page proof, and an independent audit by GPT-5.6 Sol found no flaws. This is the third major AI mathematics result in 35 days — but the first in which AI built a new mathematical object rather than searching an existing one.
The 78-Year-Old Problem Nobody Could Close
Among all spheres, only two can even carry a complex structure: S² and S⁶. The first is classical and long settled; the second has been the battleground ever since the question was posed in 1948. Generations of mathematicians split into two camps — it exists versus it cannot exist — and both sides stumbled. In 2016, Fields medalist Michael Atiyah publicly claimed a proof; it collapsed because the argument had gaps. Even the late Chinese master Shiing-Shen Chern studied the problem late in his career.
Alpöge, a junior fellow at Harvard's Society of Fellows and a postdoc at Anthropic, is nominally a number theorist and arithmetic geometer — not a complex geometer. His announcement on X greeted the result like a birth: "Welcome this beautiful new geometric object to the world."
The Construction: Building a New Object
Instead of arguing about whether a complex structure can exist, Alpöge and Claude simply built one. The construction runs in three moves:
Step 1 — the base. Fold the upper half-plane using the triangle group (3,4,∞). The quotient looks like a sphere with three marked points: one of order 3 at t=0, one of order 4 at t=1, and a cusp at t=∞.
Step 2 — attach torus fibers. Over every point except those three special points, mount a complex 2-torus — a fiber that is two-dimensional in the complex sense and four-dimensional in the real sense. The three marked points stay empty, leaving three holes punched through the sphere.
Step 3 — fill the holes. Each hole is filled by a different classical technique that happens to fit perfectly. The cusp at t=∞ uses Mumford's toroidal degeneration: fiber W is built by gluing opposite edges of a hexagonal boundary of a sextic del Pezzo surface. The points t=0 and t=1 use Kodaira's logarithmic transforms with multiplicities 3 and 4 — exactly matching the orders of the base points. When the holes close, a compact complex 3-fold X is born.
Why This Proof Holds Up
Construction alone is not enough — X must actually be S⁶. The proof's Section 7 computes the fundamental group:
π₁(X) ≅ ℤ / |12ℓ₀ − 4ℓ₁ − 3ℓ₂|
Plugging in (ℓ₀, ℓ₁, ℓ₂) = (0, 1, −1) gives |−1| = 1, the trivial group. The sphere is simply connected, so X matches S⁶ on the first criterion. The integral homology lines up as well; with the Hurewicz and Whitehead theorems, X is a homotopy 6-sphere. Smale's 1961 generalized Poincaré conjecture upgrades that to homeomorphic to S⁶. The last hurdle is smooth structure: in 1963, Kervaire and Milnor counted exotic spheres — zero in dimension 6, while dimension 7 has 28. Homeomorphic becomes diffeomorphic, and X is S⁶ itself.
This is where AI verification enters. Mathematician Qiaochu Yuan challenged GPT-5.6 Sol to find flaws in the 108 pages. After six minutes of scrutiny: nothing. After fifteen more: still nothing — the audit actually clarified the argument. Sol's verdict: if the proof stands, it could be the most important AI math result to date; purely human work of this significance might be Fields-worthy.
The Real Shift: From Finding Answers to Building Objects
The S⁶ result is the third in a 35-day sequence:
- July 20 — Alpöge and Claude Fable 5 produced a counterexample to the Jacobian conjecture, a problem open since 1939 (87 years).
- August 10 — an unreleased research version of Claude raised the proven proportion of Riemann zeta zeros on the critical line from 41.6% to 67.2%, orchestrating ~60 sub-agents, 2,400+ shell commands, and 31 million output tokens.
- August 24 — the S⁶ complex structure.
The first two results could still be explained away as an extremely powerful search over a known solution space. This one cannot. The object did not exist before; the model helped create it. Justin Curry, associate professor of mathematics and statistics at SUNY Albany, said that if the proof is correct, it is among the most remarkable AI achievements of recent memory. For 78 years the question hanging over S⁶ was does a complex structure exist? The new question is how many more are hiding there?
What This Means for AI and Scientific Discovery
Three structural changes are visible from the outside. First, a verification economy is taking shape: AI audits AI, and models like GPT-5.6 Sol become referees for frontier claims — the bottleneck is no longer generating a proof but certifying one. Second, research now has a compute budget: 31 million tokens and sixty agents for a single theorem push (the orchestration infrastructure behind this is the subject of the agent runtime war). Third, openness compounds: explicit matrices, coordinates, and gluings make proofs machine-checkable, and the open-weight push exemplified by Nvidia's $6B move into open models widens who can afford to run this kind of research.
What You Can Do Now
- Researchers: put a frontier model in the loop as an auditor — have it check proofs line by line, and publish artifacts (coordinates, matrices, gluing data) so claims are machine-checkable from day one.
- Engineers: build verification loops into agent systems: checker agents, replay, and rollback, with provenance for every step. Fifty-agent runs fail loudly when state is not auditable.
- Teams: budget compute for research the way you budget GPUs for training, and credit models as collaborators — attribution will drive who gets to publish what.
Frequently Asked Questions
Q: Did AI really solve a 78-year-old math problem?
A: Anthropic's Claude worked with Harvard mathematician Levent Alpöge to construct a complex structure on the six-dimensional sphere S⁶, a problem open since 1948. The 108-page proof passed an independent GPT-5.6 Sol audit with no flaws found — the third major AI math result in 35 days.
Q: Why is the S⁶ complex structure problem so hard?
A: Only S² and S⁶ among all spheres can possibly carry a complex structure; S² was settled long ago, leaving S⁶ as the lone battleground for 78 years. A 2016 claim by Fields medalist Michael Atiyah turned out to have gaps. Alpöge and Claude succeeded by explicitly building the object and verifying it is genuinely S⁶ — trivial fundamental group, matching homology, and no exotic spheres in dimension 6.
Q: What changes about AI's role in research?
A: Earlier AI math searched known solution spaces; the S⁶ result constructs a brand-new object, signaling a shift from AI as answer-finder to AI as object-builder, with AI-auditing-AI becoming the new verification bottleneck. Full paper: alpo.ge/s6.pdf.
