On August 1, OpenAI announced its next frontier model inside a blog post titled “Ten advances in mathematics and theoretical computer science.” The name, Astra, appears in paragraph three, described in the company’s own words as “our next major model.” There’s no release date, no pricing tier, no keynote. The reveal is technically a footnote in a research write-up.

The write-up itself is the point. An internal version of Astra produced solutions to ten open problems in math and theoretical computer science, each of which had sat without meaningful progress for at least a decade. The problems span seven fields, from high-dimensional geometry and coding theory to quantum complexity and lattice cryptography. Notable results include new upper bounds on sphere-packing density down to the Cohn-Elkies threshold, a construction establishing the existence of non-sofic groups, and a disproof of Connes’s rigidity conjecture. Every proof was formalized as a Lean certificate and pushed to a public GitHub repo, openai/ten-proofs. Total compute cost at GPT-5.6 Sol API rates: roughly $2,000.

That number is the actual announcement. Not the name.

The mathematical community’s reaction ran through the usual channel. Thomas Bloom, the University of Manchester mathematician who runs erdosproblems.com, called it “big news” on X, while rejecting the framing that AI is replacing mathematicians, a claim he argued makes little sense when the model is drawing on more than a century of accumulated theory. Noam Brown, the researcher behind the test-time reasoning technology Astra uses, hedged from the other direction: “Sadly, no Millennium Prize Problems (yet).”

Product identity is deliberately unresolved. According to The Information, Astra is a new model family built for long-running workloads that coordinate multiple agents against parts of a larger problem. It could ship as GPT-5.6 Astra, GPT-6 Astra, or just Astra, slotting into a celestial tier that already includes Terra, Luna, and Sol. Sam Altman has reportedly demoed it in Washington, D.C., and it’s expected to be the first model routed through a planned U.S. government review process requiring official approval before public release.

Read the sequencing. The math results are the credential; the regulatory review is the constraint; the name is the tell. OpenAI is signaling that the frontier model conversation is no longer a launch event, it’s a clearance process.

Sources