OpenAI's Astra cracks 10 math theorems for $2,000 in tokens—and why that matters for research

An internal iteration of OpenAI's Astra frontier model has generated machine-verified proofs for 10 long-standing mathematics and theoretical computer science problems at a cost of roughly $2,000 in API tokens. The breakthrough itself matters, but the price tag may signal something more consequential: a shift toward reasoning as a budgetable resource.

The Real Insight: Cost as Opportunity#

OpenAI hasn't yet released Astra publicly or announced its pricing, and the $2,000 figure is extrapolated from GPT-5.6 Sol API rates. That projection glosses over the enormous expenses of training and the expert labor needed to select problems and validate answers. Yet for developers and research teams, the metric opens a practical question: how much reasoning power is worth spending on a given problem?

Unlike vague claims that a model "excels at mathematics," an inference budget is something labs and startups can actually plan around. A research institution might allocate $5,000 to explore a conjecture; a pharmaceutical company might justify $50,000 to accelerate drug discovery. The calculus changes when frontier reasoning becomes a paid service rather than an exclusive tool locked inside a company's walls.

Inference Budgets Over Training Costs#

The economics of AI have historically favored the handful of organizations wealthy enough to build their own models. The training and infrastructure costs run to billions. But if smaller teams can rent access to a finished model's reasoning capabilities through an API—much as they rent cloud computing instead of operating data centers—the competitive landscape shifts.

Research labs cannot easily budget around abstract capabilities; they can budget around concrete compute costs. This reframes who gets to experiment with cutting-edge AI: not just those who can afford to build, but those who can afford to query.

What Didn't Happen#

The work doesn't include solutions to any of the Millennium Prize Problems, and human mathematicians still need to validate the proofs. OpenAI researcher Noam Brown noted that the model used relatively modest test-time compute, suggesting that larger reasoning budgets could potentially yield more discoveries. There's no guarantee that spending more tokens unlocks breakthroughs—uncertainty is inherent to research—but the framework makes serious investigation feasible for teams without frontier models of their own.