On August 1, 2026, OpenAI announced Astra, calling it their next major model family. They didn't release a product. They dropped a GitHub repo containing Lean 4 certificates formally verifying ten solutions to open problems in mathematics, some unsolved for over a decade. The standout: a construction proving non-sofic groups exist, a question Mikhail Gromov posed in 1999 that sat open for 27 years.
The estimated token cost to find all ten solutions: roughly $2,000 at Sol API rates. That's about $200 per problem spanning group theory, von Neumann algebras, quantum complexity, and lattice cryptography.
Sit with that number for a moment.
What Astra Actually Is
Astra is not a public product yet. OpenAI is positioning it as a model family built for long-horizon multi-agent work. The design is explicit: multiple agents working together on a single task for hours or days, not seconds. The system plans, tests its own output, revises, and keeps going without needing you to steer each step.
This is a different design philosophy than GPT-5.6 (Sol, Terra, Luna), which shipped in July 2026 for fast, well-framed single-turn and short-context work. Astra is aimed at the class of problem where you cannot write a good prompt because you don't know what the sub-problems are yet.
The math results are the evidence OpenAI chose to put forward first.
The Math Results and Why They're Hard to Dismiss
Ten problems across group theory, von Neumann algebras, high-dimensional geometry, quantum complexity, lattice cryptography, and extremal combinatorics. OpenAI describes each as having seen no material progress on the main result for at least a decade, and some were open far longer.
What makes this more credible than a typical benchmark claim: each solution comes with a machine-checkable Lean 4 certificate. Lean is a formal proof language where correctness is verified by a compiler. You don't have to take OpenAI's word for it. The certificates are publicly available under Apache 2.0 on OpenAI's GitHub, and any reader with a laptop can run them and see whether the proofs type-check.
Thomas Bloom, a mathematician at the University of Manchester who runs erdosproblems.com, called the results "big news" and said the constructions were bigger than the unit-distance graph counterexample OpenAI had disclosed earlier in the year.
The non-sofic group result is getting the most attention. Gromov introduced soficity in 1999, asking whether all groups can be approximated by finite symmetric groups. That question sat unanswered for 27 years. Astra produced a construction that closes it. The announcement and the Lean proof files are here.
What the $2,000 Cost Actually Means
$2,000 for ten results across multiple fields of pure mathematics is not a small number, but in context it's striking. A hard research problem in math historically requires a skilled human, often years of effort, and sometimes a lucky insight that nobody could have scheduled or budgeted.
The per-problem figure doesn't mean Astra will solve your specific hard problem at $200 a shot. Math proofs are probably the best-case showcase for formal AI reasoning, not the baseline for general work. The system had the benefit of well-specified problems with known solution criteria.
But it does establish something real: the cost of frontier AI agents on genuinely hard formal problems is in a range where an organization can think about it systematically. That's new. A year ago this sentence wouldn't have made sense.
I haven't run this at scale, and I don't know how many failed attempts or wasted compute went into those results. But even accounting for overhead, $200 per open math problem changes what's worth trying.
Formal Verification as the Trust Primitive
The Lean 4 angle is not a flex about mathematics. It's a statement about how you make AI output trustworthy on consequential tasks.
The main problem with deploying AI agents for serious work is that you can't easily verify the result. A well-written bug fix that compiles is not necessarily a correct one. A plausible architecture proposal is not necessarily a sound one. You end up trusting the model, which means trusting the training, which means hoping.
Formal verification breaks that loop for a specific class of tasks: anything expressible as a logical system with checkable proofs. Mathematical theorems, type safety, certain security properties, and some program correctness claims all qualify. When an agent produces a Lean 4 proof, the proof is correct or the compiler rejects it. The trust goes to the checker, not the model.
I don't know how Astra handles tasks where formal verification doesn't apply (which is most of software engineering). But if long-horizon agent systems converge on "produce a formally verifiable artifact" as the output standard for high-stakes work, that's a significant architectural shift. Almost nobody is building pipelines that end in a Lean or Coq certificate today. That may start to change.
The Government Review Requirement
OpenAI is framing Astra as the first model family to go through a formal U.S. government pre-release review. Sam Altman was on Capitol Hill in early August, meeting with members of the intelligence committee including Sen. Mark Warner. The specifics of what the review actually requires are still vague.
For builders, the practical implication is simple: there will be a lag between "OpenAI has this capability" and "you can call it from an API." That gap could be months. Factor that into your roadmap if you're betting on Astra-tier capabilities in production.
What to Watch
The math proofs are public now. If you want to see what Astra's output actually looks like at the limit of its current capabilities, the Lean 4 files are on OpenAI's GitHub under Apache 2.0. That's more useful than any model card claim.
What I'm actually watching: how the multi-agent architecture gets exposed as an API primitive. If Astra ships with tooling for defining long-horizon agent workflows and specifying formal output schemas, that's the part that will matter most for teams building real systems. The math is a proof of concept. The architecture underneath it determines whether any of this translates to the work most engineering teams actually need to do.
Comments
Post a Comment