"our next major model" — OpenAI.
What Astra reportedly achieved
OpenAI revealed an unreleased model, Astra, after an internal version produced ten significant advances in mathematics and theoretical computer science. According to OpenAI, the solved problems had seen no progress on their central results for at least a decade and, in most cases, much longer. The company said its internal research covered a wide range of technical fields, explicitly naming high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, lattice cryptography, and extremal combinatorics.
Named mathematical results and verification steps
OpenAI highlighted several concrete outcomes the internal model produced: the existence of non-sofic groups, a disproof of Connes’s rigidity conjecture, new bounds for high-dimensional sphere packing, and results resolving several problems posed by mathematician Paul Erdős. Human researchers used the same model to prepare the arguments as manuscripts, and Astra then formalized every argument as a Lean certificate, allowing the proofs to be checked using the mathematical verification system Lean.
Built by Nubivance.
OSINTSights' secure edge-first architecture, AI content pipeline, and serverless ops are designed by Nubivance. We do this for clients too.
Talk to us →Cost and compute context reported by OpenAI
OpenAI provided a specific compute-context metric: "The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates," the company noted. The post framed Astra as a model designed for long-running workloads and collaborative workflows, saying the model allows AI agents to collaborate on different parts of a larger problem.
Release options and independent confirmation
The Information independently confirmed OpenAI is working on Astra, describing it as a new model family built for long-running workloads. Reporting compiled by BleepingComputer noted OpenAI had not yet decided whether Astra will be released under the GPT-5.7 or GPT-6 name — or under another designation. The coverage also flagged that Astra could be subject to "Anthropic-like policies" in which one version is released to consumers while a more powerful variant would require special approval.
What this means for scientific researchers, policymakers, and enterprises
- Scientific researchers and mathematicians: The production of Lean certificates and the use of a formal verification system mean researchers can inspect machine-generated proofs in a machine-checkable form; the source specifically says Astra formalized every argument as Lean certificates and enabled checking via the verification system.
- Policymakers and regulators: The source raises a release-path question — Astra "could be subject to Anthropic-like policies" — signaling decisions about consumer access versus restricted approval could determine who sees more powerful variants.
- Enterprises and procurement leaders: Reported undecided naming and release plans (GPT-5.7, GPT-6, or another name) and the $2,000 token-cost figure at Sol API rates are concrete data points buyers can use when assessing potential acquisition or research partnerships.
OpenAI framed Astra as a significant step toward applying large models to complex, long-running scientific problems and noted the internal version produced advances in multiple deep subfields of mathematics and theoretical computer science. Independent reporting confirmed work is underway and flagged open questions around naming and controlled release. The next concrete milestones the company faces, per the reporting, are whether to release Astra publicly and under what name or access controls — decisions the source says remain undecided.
