Frontier models produced counterexamples to a 1967 conjecture, at $116.28 total. Math research just got an infrastructure price

AI did something mathematicians had treated as a decades-long hunt, and the experiment's cost is the part worth budgeting.

Per the weekly coverage, frontier models including GPT-6 Astra Pro, Gemini 4, Claude Opus 5.5, and Claude Sonnet 5.5 produced counterexamples to a conjecture that had stood since 1967.

The result landed inside a 50-problem batch run, and the run cost $116.28 total.

That averages under $3 per problem, per the same coverage.

Careful in one place, since sourcing on this story runs thin: the conjecture identification sits in recap coverage rather than a paper page, so replication is the thing to track. The number is the part that carries.

What a $3 budget line changes

Follow the money into the research structure it prices.

Counterexample search is a class of work mathematics treats as expensive: a proof assistant plus domain effort, usually over weeks. A $116 batch run prices a search that humans schedule over a semester.

Under the hood, the format is the mechanism. A model arms the search against problem with a second model, and the batch cost splits across the model set.

Since models work through every one of them for under $3, failure is cheap, and the error cost of running again is contained.

Precedent in research price: model-assisted work ran longer at per-task costs, and this month already carried compute-for-science stories, with Astra publishing math work on prime gaps over a launch.

The unresolved edge is verification itself. Counterexamples to a conjecture need proofs, and machine finding is no substitute for a human-checkable derivation.

I get the sense the research market prices this one through what the Astra math launch showed. Speculative search has a real bill.

If you run a lab or a sourcing team, the useful read is that the under-$3-per-problem number now prices a whole research class that was reservation-bound.

Source