AI Pulse · Frontier Labs Watch

OpenAI just published a 249-page report claiming its unreleased Astra model solved ten open problems in math, quantum complexity, and theoretical computer science, including what it says is the first explicit example of a non-sofic group. Each proof is machine-checkable in Lean, and the company says the compute cost was around $2,000. The catch? No one outside OpenAI can run the model, and no external researchers have verified the claims. OpenAI graded its own homework.

The timing matters because frontier labs are racing to show progress on reasoning and formal proof. A model that can tackle decade-old open problems for pocket change would be a meaningful shift. But the details are messy. At least one result being shared widely, the non-sofic group claim, is being called a joke or parody by some accounts, including Grok’s own X feed. The underlying report is real, but individual technical claims are contested or unclear. The original report is live, but treat specifics with caution until outside mathematicians weigh in.

For business owners, this is less about the proofs themselves and more about what happens when a lab releases big claims without external review. If Astra can do what OpenAI says, it changes the cost structure for formal verification, code correctness, and technical research. If it can’t, or if the claims are overstated, it’s a reminder that model capability announcements need independent confirmation before you build on them. This is exactly the kind of signal we track in the Omni Command Centre, so you know which breakthroughs are real and which are still developing. Right now, Astra is a black box with bold claims. Wait for the math community to respond before treating this as settled.