OpenAI's next model proved its math without asking anyone to trust it. Structural engineering AI still runs entirely on trust.
OpenAI says an internal version of its next model, Astra, solved ten math and computer-science problems that sat unsolved for a decade or more — and backed every result with a proof a machine can verify, not a narrative you have to take on faith. That's the exact capability AI tools touching structural calculations still don't have.
OpenAI says an internal version of its next model, code-named Astra, just solved ten math and computer-science problems that had gone unsolved for a decade or more — and it proved every one of them in a form a computer can check line by line, without asking anyone to trust the model. That verification trick, more than the math itself, is worth a construction reader's attention: it's the exact capability that AI tools touching stamped structural calculations still don't have.
What did Astra actually do?
On August 1, OpenAI published a 249-page manuscript of results from Astra alongside written reasoning walkthroughs and formal proof certificates on GitHub. The headline result is an explicit construction of a non-sofic group — a question open since Mikhail Gromov introduced soficity in 1999 — plus a disproof of Connes's rigidity conjecture, a proof of Ehrhart's volume conjecture, improved sphere-packing and coding-theory bounds, and counterexamples resolving several problems on Paul Erdős's famous open-problem list. Astra itself isn't public. None of the ten results has been through peer review yet, and OpenAI has given no release date for the model.
Why does "machine-checkable" matter more than "AI solved math"?
AI models producing a plausible-looking answer is not news. What's new is that each result shipped with a Lean 4 certificate, and Lean's compiler returns one of two verdicts: the proof holds, or it doesn't. The GitHub repository reports a "sorry" count of zero across all ten — meaning no step in any proof was left unproven or hand-waved. You don't have to trust Astra's confidence, its tone, or OpenAI's summary. You run the certificate through an independent checker and it either compiles or it fails.
That's also where the limit sits. Lean confirms the internal logic is airtight — it doesn't confirm that the formal statement being checked actually says what the original open problem asked. A mathematician still has to verify that translation, and judge whether the result matters. Formal verification removes one kind of trust and leaves another exactly where it was.
Does this fix AI's trust problem on stamped structural calcs?
Not yet, and the gap is specific. As Structure magazine has documented, AI tools already in use in structural engineering practice sit upstream of the stamp — handling the 15 to 20 design iterations a team runs through during schematic design and design development while searching for the right structural system. The final calculation and the stamp on it are still entirely the licensed engineer's, and no building department treats "the AI said it's safe" as a basis for permit approval.
The reason Astra's trick doesn't transfer directly: building codes are natural-language documents with genuine ambiguity, not formal systems with a compiler. ASCE 7 load combinations and IBC provisions don't have a Lean equivalent that could take a load calculation and return a binary "this complies" or "this doesn't." Nobody has built a formally verifiable encoding of a building code the way OpenAI just built one for a set of open math problems — which means the trust question construction AI vendors are actually solving today is still "does a licensed engineer agree with this," not "does an independent checker confirm this."
What to do with this now
- Keep AI structural tools where they already live — early-stage exploration and iteration during schematic/DD, not final calc sign-off. That hasn't changed.
- Ask every vendor touching engineering output the same question: does this come with any independent, machine-verifiable check, or is it the model's narrative? Today it's almost always the latter, and budgeting the same PE review hours you do now is the honest call until that changes.
- Watch for a "Lean for the building code" — a formal, checkable encoding of load and life-safety provisions. That's the actual unlock this story points toward. Nobody's shipped it.
We wrote last week about a fabricated AI-generated court citation that made it into a live construction lawsuit — a case where nothing checked the AI's output before it hit a filing. Astra's proofs are what it looks like when something does. Structural engineering doesn't have that yet, and that gap, not the math, is the story.
Forward this to whoever on your team is evaluating an AI tool that touches engineering calculations.
- What did OpenAI's Astra model actually solve?
- OpenAI published results from an internal, not-yet-released version of Astra — the model it's calling its next major release — on August 1, 2026: ten problems in math and theoretical computer science open for a decade or more, including an explicit construction of a non-sofic group (open since 1999), a disproof of Connes's rigidity conjecture, a proof of Ehrhart's volume conjecture, and counterexamples resolving several Erdős problems in extremal graph theory.
- What makes these proofs different from a normal AI answer?
- Every result shipped with a Lean 4 formal proof certificate on GitHub, not just a written explanation. Lean's compiler returns a binary verdict — the proof checks out or it doesn't — so verifying the result doesn't require trusting Astra's reasoning or taking OpenAI's word for it. Anyone with the Lean compiler can run the check themselves.
- Does this mean AI can now be trusted for stamped structural calculations?
- No. Building codes like ASCE 7 and the IBC aren't written in a formal, machine-checkable language the way pure mathematics is, so there's no equivalent compiler to verify a load calculation against them. No building department accepts an AI's output as the basis for a permit, and the licensed engineer of record still carries full legal liability for the final stamped set.
- What should a GC actually do with AI tools that touch engineering calculations?
- Use them the way structural engineers already do: as an exploration aid for early design iteration, not a substitute for the engineer's final review. Ask any vendor whether their output comes with independent verification or is just the model's confident narrative — today, it's almost always the latter, and the PE review hours don't go away because of it.