OpenAI Announces Astra — the Model That Solved Ten Long-Standing Math Problems

OpenAI has revealed a new model codenamed Astra, built for hard problems that require long computation cycles. In a research post, Astra is described as “our next major model,” and the announcement was prompted by ten results that an internal version of the system produced in mathematics and theoretical computer science.

According to OpenAI, these are problems whose central components had seen no progress for at least a decade — and in most cases considerably longer. The work spanned a wide range of areas: high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, lattice-based cryptography and extremal combinatorics.

What exactly was solved

Among the examples cited are the existence of non-sofic groups, a refutation of Connes' rigidity conjecture, new bounds for sphere packing in high-dimensional spaces, and solutions to several problems posed by the mathematician Paul Erdős. OpenAI also singles out rapid progress in the natural sciences.

A curious detail for those who track the economics of computation: by the company's own reckoning, the total number of tokens required to find solutions to all these problems would have cost roughly $2,000 at Sol API rates. In other words, the price of a breakthrough — if the framing is to be believed — is comparable to a modest purchase.

The workflow looked like this: human researchers, using the same model, wrote up the arguments as manuscripts, after which Astra formalized each proof as a certificate in the Lean language — allowing the reasoning to be checked in a system for the automated verification of mathematical proofs.

Under what name will it ship

The outlet The Information independently confirmed that OpenAI is indeed working on an Astra family aimed at long-running workloads. The model is described as powerful enough for AI agents to work together on different parts of a single large problem.

A final name has not yet been chosen: according to BleepingComputer, the release could come out under the label GPT-5.7, GPT-6 or some other name. A scenario in the spirit of Anthropic's policy is also possible, in which users are given one version while access to a more powerful variant requires separate approval.

For now Astra remains unreleased, and all the loud claims come from the company itself — the post offers no independent verification of the listed mathematical results by outside specialists.

Source: Bleeping Computer