Astra Model Solves 10 Open Math Problems With Lean 4 Proofs

On August 1, 2026, OpenAI announced Astra, a new model family designed for long-running workloads that has resolved 10 open problems in advanced mathematics and theoretical computer science. The release includes a 249-page manuscript collection detailing breakthroughs in fields where researchers have seen no significant progress for up to three decades.

Architecture for Continuous Research

Astra represents a shift from synchronous chat models to continuous, multi-step systems. OpenAI describes the model as a research collaborator that manages specialized agents working continuously over hours or days. Rather than generating a single immediate response, Astra delegates sub-tasks to subordinate agents that explore parallel branches of a mathematical proof.

Running models in this continuous state changes the economics of inference. The compute expenditure to solve all 10 problems required an estimated $2,000 at current GPT-5.6 Sol API token rates. This expenditure highlights the shift from cost-per-query to cost-per-result when dealing with autonomous reasoning workloads.

Every mathematical claim produced by Astra is backed by a Lean 4 certificate. Lean acts as a formal proof assistant, compiling mathematical logic into machine-checkable steps to eliminate verification ambiguity. The model does not just output a text-based proof; it generates the strict syntax required for algorithmic validation.

Verified Mathematical Results

The 10 resolved problems span multiple advanced disciplines, including arithmetic circuit complexity, quantum complexity, and lattice cryptography. The most notable results include long-standing conjectures.

Field	Result	Verification Status
Operator Algebras	Disproof of Connes’s rigidity conjecture	Lean 4 Certified
High-Dimensional Geometry	Improved upper bound on sphere-packing density	Lean 4 Certified
Group Theory	Proved the existence of non-sofic groups	Lean 4 Certified
Combinatorics	Resolved 3 Paul Erdős catalog problems	Lean 4 Certified

In May 2026, an early version of the Astra architecture disproved the 80-year-old Erdős unit distance conjecture. The newly published manuscript builds on that foundation. The improvements to the general upper bound on sphere-packing density in high dimensions represent the first mathematical progress on that specific problem since 1978.

Replication and Regulatory Review

Replication of these results indicates the methodology relies heavily on modern search-and-verify pipelines rather than purely upon parametric knowledge. On August 2, Anthropic researcher Levent Alpöge replicated approximately half of the mathematical results in 24 hours using Anthropic’s Fable 5 model. This rapid replication suggests the core breakthrough involves identifying which specific mathematical conjectures are highly susceptible to targeted multi-agent coordination patterns.

Astra will be the first OpenAI model family subjected to a new U.S. government regulatory review process requiring official approval before public release. Following the 2026 Latin celestial naming convention of Terra, Luna, and Sol, Astra serves as the foundation for the next major API iteration, heavily speculated to be GPT-5.7 or GPT-6.

If you build long-running autonomous systems, the Lean 4 integration demonstrates the necessity of deterministic verification layers. To reliably evaluate and test AI agents operating over long time horizons, route complex logic outputs through strict programmatic verifiers rather than relying on the model to self-correct its own reasoning.