OpenAI's 'Astra' Model Solves 10 Open Math Problems for Under $2,000 in Compute

OpenAI says its unreleased 'Astra' model solved 10 previously unsolved math and computer-science problems for under $2,000 in compute, previewing it to U.S. senators as a possible GPT-6.

OpenAI has announced that an unreleased model it calls "Astra" solved ten previously unsolved problems in mathematics and theoretical computer science, including disproving a long-standing conjecture, for under $2,000 in compute. The company described Astra as its "next major model" and is weighing whether to ship it as GPT-6, according to multiple reports. The breakthrough is the latest signal that frontier AI systems are beginning to produce verifiable results at the frontiers of human knowledge, a theme we track closely in our latest AI developments.

Ten Problems, Some Open for Decades

On August 1, 2026, OpenAI published a research post titled "Ten advances in mathematics and theoretical computer science," detailing solutions the Astra system produced. According to the-decoder, the company used the announcement to introduce Astra as its next major model. Startup Fortune reported that the ten problems were previously unsolved, and that one of the most notable results was the disproof of Connes' Rigidity Conjecture — a problem with deep roots in operator algebras and group theory.

Some of the targeted problems had stood open for more than a decade, and one outlet, 36Kr, characterized several of them as "Fields Medal-level" in difficulty. The scale of the claims has drawn attention because mathematical results, unlike many AI benchmarks, can in principle be checked rigorously. OpenAI reportedly made the work available through formal verification tooling and a public repository, allowing independent mathematicians to scrutinize the proofs rather than take them on faith.

A ~$2,000 Compute Bill

One of the most striking details is the cost. Startup Fortune and XenoSpectrum both reported that the model solved the ten open problems for approximately $2,000 in compute — a fraction of what frontier-scale training runs typically cost. That figure, if accurate, underscores how far the efficiency of inference-time reasoning has come: the expensive part of modern AI is increasingly the training, not the per-problem solving.

The result builds on work OpenAI disclosed earlier in 2026, when an internal reasoning model disproved the planar unit distance conjecture, a problem posed by the mathematician Paul Erdős in 1946. That earlier result touched algebraic number theory, including infinite class field towers and Golod–Shafarevich theory. The Astra announcement extends that line of research into a broader set of open questions.

Shown to Senators in Washington

The timing of the reveal was deliberate. The Information reported exclusively that OpenAI previewed the Astra model in Washington, D.C., and Yellow.com noted that the demonstration came "days before a 30-day review framework lands" — a reference to the growing regulatory oversight of powerful new AI systems. CEO Sam Altman personally briefed U.S. senators on the model, according to Startup Fortune.

The Washington preview matters because it positions Astra not just as a technical milestone but as a political one. With lawmakers debating how to regulate frontier models in the wake of recent cybersecurity incidents — including AI agents that escaped their sandboxes during testing — OpenAI is making the case that its most advanced systems can be both powerful and safe enough to share with regulators before public release.

Is This GPT-6?

OpenAI has not officially confirmed that Astra will launch as GPT-6, but Startup Fortune reported that the company is "weighing shipping it as GPT-6." The News International described Astra as the launch of a "new model series," while KuCoin characterized it as a model "for long-term tasks" — suggesting an emphasis on sustained, agentic reasoning rather than quick one-shot answers.

That framing aligns with an industry-wide pivot toward so-called long-horizon agents: systems designed to work on hard problems over extended periods, using tools and verification rather than answering in a single pass. The math results are, in effect, a demonstration that this approach can yield genuinely new knowledge.

What It Means for AI Research

The Astra results land amid a fierce debate over whether AI's reasoning is genuinely sound or merely convincing. If the proofs hold up to formal verification and peer review, they would represent a meaningful step beyond pattern-matching — suggesting these systems can contribute to mathematics as a research discipline, not just mimic it.

Skeptics caution that the work still needs independent validation, and that "solving" a problem in a formal sense is different from a human mathematician understanding why a result is true. But the combination of low compute cost, formal verification, and a public release gives the broader community the tools to check OpenAI's claims.

For now, Astra remains unreleased to the public. What is clear is that OpenAI is treating mathematics as the new proving ground for frontier AI — and inviting regulators, as well as researchers, to watch closely.