OpenAI previewed its next major model family — reported as "Astra" — in early August 2026 by doing something concrete: pushing machine-checkable proofs of ten previously unsolved math and theoretical-computer-science problems to a public repository. The problems span group theory, coding theory, quantum complexity, and high-dimensional geometry, each reportedly open for at least a decade.
How it did it
Astra is described as a multi-agent system: it breaks a hard problem into pieces and coordinates a team of sub-agents working over hours or days, rather than answering in one shot. This is test-time compute taken to an extreme — spend a lot of inference, orchestrated across agents, to crack problems a single forward pass never could.
The headline isn't "AI did math." It's how — decompose, delegate to sub-agents, grind for hours, then verify the proof by machine. That's a template, not a stunt.
The $2,000 detail
Reporting put the compute cost of solving all ten at roughly $2,000. If accurate, that reframes the economics of research: certain hard problems may be attackable for the price of a laptop, given the right orchestration. That's a striking claim worth watching as independent verification comes in.
Confirmed vs. reported
The proofs being machine-checkable is the strong part — those can be independently verified. Astra's naming, release date, and whether it ships as GPT-6, a GPT-5 point release, or standalone are unsettled, with no public pricing. Treat the capability demo as real and the product details as unconfirmed.
Why it matters
Astra is a signal of where frontier work is heading: not just bigger models, but systems of agents spending serious inference on long-horizon problems, with verifiable outputs. If a $2,000 run can settle a decade-old problem, expect a wave of AI-assisted research — and hard questions about how we credit and verify it.
