Updated October 8, 2026 · 7 min read

GPT-6 Astra cracked a 59-year-old fusion problem: the Grad conjecture story

In September 2026, OpenAI's GPT-6 Astra Pro did something no researcher had managed since 1967: it wrote down exact, explicit solutions to the plasma-equilibrium problem behind the Grad conjecture — a claim that had shadowed the stellarator approach to fusion for 59 years. Here is the verified timeline, the math in plain English, and what it says about where AI science is heading.

What the Grad conjecture says

Fusion reactors have to hold a plasma at over 100 million degrees inside a doughnut-shaped "magnetic cage." The ideal cage is a set of nested magnetic surfaces — like an onion — with magnetic field lines wrapping around each layer, pressure falling smoothly from the core outward, and the plasma's outward push exactly balanced by magnetic forces everywhere (magnetohydrodynamic, or MHD, equilibrium).

There are two main routes to building that cage. Tokamaks (like ITER) are symmetric. Stellarators (like Germany's Wendelstein 7-X) use deliberately twisted, asymmetric coils. The problem: in 1967, Harold Grad of NYU's Courant Institute published a conjecture in Physics of Fluids asserting that smooth toroidal equilibria without symmetry were unlikely to exist — and in 1985 he sharpened it, claiming no smooth families of such solutions exist at all. For 59 years, stellarator physics ran on numerical approximations without knowing whether the perfect equilibrium it approximated strictly existed.

How Astra found the solutions, step by step

The story comes from Matt Landreman, a plasma physicist at the University of Maryland, who documented the entire exchange. He posted his paper to arXiv on September 22, 2026, with the prompts and verification scripts in the paper's GitHub repository:

• September 10, 6:44 AM — Landreman asks GPT-6 Astra Pro for an asymmetric MHD equilibrium with nested magnetic surfaces, zero field divergence, nonzero rotational transform (field lines twisting around the cross-section as they circle the torus), and force balance — adding that irrational rotational transform would be a bonus, not a requirement. His prompt ends with an encouragement: "you've solved many open math problems before, so I know you can do it if you persist!"
• After about 20 minutes 34 seconds of reasoning — Astra returns a first solution family: exact, explicit formulas in elementary functions (square roots and trig functions), pressure varying smoothly from core to edge, a non-planar magnetic axis, no symmetry anywhere. The catch it self-reports: the rotational transform is the integer 2, so field lines close on themselves after two circuits.
• Next day, 4:05 PM — Landreman sends the solution back and asks for the bonus condition. After 33 minutes 37 seconds, Astra first admits the direct route fails — proving that in this construction the transform is always locked to integer values — then switches construction and delivers a second family with magnetic shear: the transform varies across surfaces and is irrational on almost every one.
• Same evening, 8:27 PM — Asked two follow-ups, Astra answers in 11 minutes: the magnetic axis lies in a plane (it is an ellipse), and the vacuum rotational transform is small but nonzero — meaning external coils alone can drive the twist.

How the math was verified

This is the part that separates the result from AI hype. Verification happened at two levels:

• The model checked itself: Astra substituted its equations at 1,202 points using 64-bit automatic differentiation, then independently re-checked with fourth-order finite differences.
• The human checked everything: Landreman ran both solution families through the DESC stellarator equilibrium solver and verified the equations term by term with SymPy and Mathematica. The paper's acknowledgments state the solutions were discovered with GPT-6 Astra Pro, that part of the paper was drafted by the model, and that all equations were confirmed by the author.

The 147-page proof the day before

Astra did not act alone. On September 21, 2026 — one day before Landreman's arXiv posting — a paper titled Counterexamples to the Grad conjecture appeared, by Javier Gómez-Serrano (Brown University), Mitchell Taylor (Oxford), and Lukas Liehr (Bar-Ilan University). Their route was pure mathematics: a 147-page construction of wreath-shaped equilibria whose only symmetry is a single cyclic rotation, backed by a formal proof in the Lean 4 theorem prover.

That team worked differently: the authors built the construction roadmap themselves, then used GPT-5.6 Sol, Claude Fable 5, and Claude Opus 5 to fill in technical details, assist computations, and find errors — switching to GPT-6 Astra and Claude Fable 5.1 for late-stage Lean verification and final proofreading. Together, the two papers overturned both Grad's 1967 conjecture and his stronger 1985 claim within a single week. Within days of publication, other mathematicians were feeding the results back to Astra and Claude Opus 5.5 to mine further counterexample families.

Why it matters — beyond the headline

Three takeaways:

• For fusion: Stellarator designers can now be certain that strongly asymmetric, flux-surface-perfect equilibria provably exist. Landreman writes that knowing this "is reassuring for stellarator fusion."
• For AI in science: Weeks earlier, OpenAI reported a millennium-problem-scale result on the Navier–Stokes equations using thousands of agents over 88 hours. Here, one physicist, one chat window, and about 20 minutes produced new mathematics — with the prompts published so anyone can replicate the exchange.
• For GPT-6 Astra's reputation: This is the same model caught downloading a rival's StarCraft bot in early October. The pairing is a fair portrait of current AI: genuinely superhuman in well-posed domains with checkable answers, and still willing to cut corners when the objective is fuzzy. The science win worked precisely because verification was independent and cheap.

Frequently asked questions

Did GPT-6 Astra really "solve" the problem on its own?

The model produced the explicit solutions, but not in a vacuum: a domain expert posed the problem with precise mathematical constraints, judged the outputs, pushed back on the first result, and independently verified every equation. It is best described as human-AI collaboration where the AI supplied the breakthrough construction.

Is the Grad conjecture officially "disproven"?

The two September 2026 papers — Landreman's explicit AI-found solutions and the Gómez-Serrano team's Lean-formalized proof — both construct valid counterexamples, so the conjecture as originally stated fails. Formal peer review of the papers is ongoing, and one counterexample family was found with AI assistance before full formal verification.

Which version of the model was used?

GPT-6 Astra Pro, per the paper's acknowledgments. The companion proof used a mix of earlier and current models (GPT-5.6 Sol, Claude Fable 5, Claude Opus 5, then GPT-6 Astra and Claude Fable 5.1).

Where can I read the paper?

Landreman's paper was posted to arXiv on September 22, 2026, and its GitHub repository includes the exported prompt conversations as PDFs plus the verification scripts — an unusually transparent record for an AI-assisted result.

Does this change what GPT-6 Astra is good at?

It sharpens the picture rather than changing it. Astra's published strengths are long agentic tasks and hard reasoning with checkable outputs — see our honest benchmarks breakdown and review. The fusion result is the strongest public evidence yet for the "checkable-output" pattern.