GPT-5.6 proved a 50-year math problem — you can't check it

On July 10, OpenAI said GPT-5.6 Sol Ultra generated a full proof of the Cycle Double Cover Conjecture — a graph-theory problem open for roughly 50 years — using 64 subagents running in parallel, in under an hour. That's a genuinely impressive result, and it's also a perfect illustration of the trap every business wiring AI into its work is walking into: the output looks authoritative, and nobody can show you how it got there.

What actually happened

Per The Decoder, OpenAI published the proof as a PDF and released the exact prompt used to produce it. The setup was aggressive: the prompt assumed a proof existed, banned internet searches, rejected partial results, ran adversarial checker agents to hunt for errors, and told the model to compute for at least eight hours before giving up. The result is about three pages of elementary graph theory. Thomas Bloom, a mathematician at the University of Manchester, gave an early read: "a very nice proof," short and elementary, that "could have been discovered in the 1980s."

Then the caveats. Full verification by the math community is still pending. The proof carries no citations to the prior work it leans on — Bloom flagged a 1983 paper whose ideas underpin the solution. The conjecture has attracted several "proofs" over the years that were later found to have gaps. And Ultra mode's subagents leave no inspectable transcript: you get one opaque answer with no record of where they disagreed, hit dead ends, or resolved conflicts.

Why it matters for your business

You are not proving graph theory. But you are, increasingly, handing AI agents work where the output is confident, formatted, and hard to check — pricing decisions, inventory reconciliations, drafted contracts, customer replies, code that ships. A three-page proof from a frontier model that its own maker can't fully verify is the same failure mode as an invoice-matching agent that returns a clean number you can't trace. Fluent and correct are different things, and the more capable the model, the more the difference hides.

The move isn't to distrust AI. It's to make verification a step you own, not a step you hope the model did. Give agents deterministic tools for anything that has a right answer — a real calculator, a database query, a schema validator — instead of trusting the model's prose. Keep a human in the loop on anything expensive or irreversible. And favor systems that show their work: a log you can read beats a "trust me" result, every time. If you can't check it, you don't ship it.

Sources: The Decoder.