OpenAI named its next major model in the third paragraph of a blog post about mathematics, which is a peculiar way to launch anything. Ten advances in mathematics and theoretical computer science, published on 1 August, credits ten new results to "an internal version of Astra, our next major model" and puts the compute at roughly $2,000 at Sol API rates.

The aggregator summaries that followed got one thing flatly wrong. Astra is not a candidate name competing with "Sol and Terra". Those already exist: Sol, Terra and Luna are GPT-5.6 variants, and Astra is Latin for stars, so the new model joins the same celestial scheme as a class alongside them, built for long-running work. What The Information reports as genuinely undecided is GPT-6 versus GPT-5.7 versus something else.

Gary Marcus treats that indecision as evidence, and it's decent evidence: a company sitting on a quantum leap doesn't agonise over the version number. His stronger card is Levent Alpöge's partial replication, done quickly with Anthropic's Fable, which puts roughly half the ten results within reach of a model that already ships. Ten proofs for $2,000 cuts both ways.

The launch date is nobody's announcement. On 6 August a well-sourced leaker claimed next week, naming a dogfood checkpoint called "mewfour" as the release candidate and the largest pretrain since GPT-4.5.

Washington has seen the model, which is where the security-review line comes from. Altman demoed it to officials, and Astra is expected to be among the first submitted under the administration's planned federal pre-release framework. Nobody has issued a verdict, and the framework isn't finished.

Sources: