OpenAI Unveils Astra — a Model That Solved Ten Decades-Old Math Problems, and Each Proof Cost $2,000 in Tokens

August 2, 2026. OpenAI published a research post that redefines what an LLM can accomplish. Ten mathematical problems — some of which had seen no progress on their central results for over a decade — have been solved by Astra, an unreleased model the company describes as “our next major model.” The total research cost: $2,000 in tokens through the Sol API.

This is not brute-force demonstration. Every proof was formalized in Lean, the mathematical proof verifier, making each result mechanically checkable. Astra did not “hallucinate” ten theorems — it produced ten Lean certificates that the mathematical community can inspect line by line.

What Astra actually solved

OpenAI targeted problems spanning seven distinct domains: high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. None of these problems had seen progress on their central results for at least a decade.

Among the announced results:

• The existence of non-sofic groups, an open question in group theory since Gromov’s work in 1999.
• A disproof of Connes’s rigidity conjecture, a pillar of noncommutative geometry formulated by Alain Connes in 1976.
• New bounds for high-dimensional sphere packing, a classical problem in combinatorial geometry popularized by Paul Erdős.
• Multiple results resolving conjectures explicitly posed by Erdős himself.

The Information independently confirmed that OpenAI is working on Astra, a model family built for long-running workloads — tasks requiring hours or even days of continuous computation. According to OpenAI, Astra allows multiple AI agents to collaborate on different parts of a larger problem, with each agent handling a subset before an orchestrator consolidates the results.

BleepingComputer reports that OpenAI has not yet decided whether the model will ship as GPT-5.7, GPT-6, or under an entirely new brand name.

Why this is different from everything before it

Since GPT-4, every new OpenAI model has primarily improved along three axes: latency, cost per token, and context length. GPT-5 brought native tool use and streaming. GPT-5.6 lowered costs and improved efficiency. But none of these models had demonstrated long-horizon reasoning comparable to a professional mathematician — the ability to pursue a line of thought across hundreds of thousands of tokens without drifting.

The announced cost — $2,000 to solve ten research-level problems — is a metric that will obsess laboratories. For comparison, a math postdoc costs a university lab between $60,000 and $100,000 per year. Even assuming a human researcher would have solved these ten problems in three years (an extremely optimistic assumption), the cost ratio is 1:100 in favor of the model.

The fact that every proof is formalized in Lean changes the nature of the debate. It is no longer about “trusting LLM output” but about mechanically verifying every reasoning step. This coupling — generative AI + formal verification — makes Astra a fundamentally different object from general-purpose chatbots.

The specter of tiered release

BleepingComputer reports that OpenAI is considering a two-tier release policy, similar to Anthropic’s approach with Claude: a “consumer” version with capped capabilities, and a “research” version subject to special approval for institutional customers.

This approach would be unprecedented for OpenAI, which has historically shipped a single model per generation. If confirmed, it would formalize a market segmentation of AI between:

• Conversational models (GPT-5, Claude) for general public and enterprise use.
• Long-reasoning models (Astra, potentially Gemini Deep Research) reserved for scientific research, vulnerability discovery, and cryptanalysis.

This segmentation raises a fundamental question: if Astra can solve mathematical conjectures fifty years old, can it also break cryptographic primitives designed in an era when this type of computation did not exist? OpenAI has not commented on this point.

What this means for CISOs and DevSecOps teams

OpenAI’s timeline suggests Astra is not a lab prototype but a product in advanced development. The security implications are direct:

• Proof automation and code auditing. A model capable of formalizing proofs in Lean can, with the same mechanisms, formally prove security properties on critical code. Cryptographic implementations, kernels, and authentication protocols could be verified automatically.
• Vulnerability discovery. If Astra can explore a mathematical search space autonomously, the translation to vulnerability research in binary or source code is straightforward. A fuzzer augmented by a reasoner of this caliber shifts the attacker-defender balance.
• Software supply chain verification. Lean formalization of critical dependencies becomes economically viable at $2,000 per target. For a dependency like OpenSSL, this is a negligible investment.

Verdict

Astra is not a faster GPT. It is a new category of AI model — the long-term automated reasoner. If OpenAI delivers on its cost and verifiability promises, every research lab, every cybersecurity agency, and every critical software vendor will need to integrate this type of tool into their analysis pipeline before 2027.

For CISOs: start identifying the software components in your infrastructure where a formal security proof would have the highest impact today. The question is no longer whether these models are coming, but who will use them first — your teams or your adversaries.

References

1. OpenAI teases Astra, its next major AI model, after it solves 10 long-standing math problems — BleepingComputer, August 2, 2026
2. OpenAI Research — Astra: our next major model (internal blog post), August 2026
3. The Information — OpenAI develops Astra model family for long-running workloads, August 2026