SEC Stops Waiting for Congress: 'Regulation Crypto' Vote Set for August 14
An AI model solves 10 unsolved math problems for $2,000 — and the company that built it had to pause its own development for safety reasons.
Three Key Facts
1. Ten open problems solved, some unresolved for 27 years
On August 1, 2026, OpenAI announced that Astra, an internal unreleased next-generation model, successfully solved 10 open problems in mathematics and theoretical computer science. One of these — the explicit construction of a "non-sofic group" — had been open since 1999, stumping mathematicians for 27 years. OpenAI published machine-checkable Lean proof files on GitHub alongside a 249-page manuscript, enabling independent verification.
2. Total compute cost: approximately $2,000
The estimated computational cost for finding all ten solutions was roughly $2,000. To put that in perspective, a PhD student's lab electricity bill for a year likely dwarfs that figure. The breakthrough here isn't just about AI intelligence — it's about "verification-driven search" achieving terrifying cost-efficiency in formal mathematics.
3. Security framework triggered "critical" threshold, development paused
Astra's capabilities in agentic coding and cybersecurity crossed the "critical" threshold in OpenAI's Preparedness Framework. The company paused certain internal development activities and imposed stricter safety controls. A model so capable that its own creator had to hit the brakes — not sci-fi, just August 2026.
WangDou's Take
$2,000. Twenty-seven years. Put those two numbers side by side and the chilling part isn't how smart AI is — it's that something humans couldn't crack in 27 years got brute-forced for the price of a cheap laptop. Anyone still earning a living on "I'm smarter than the machine" needs to start rethinking their retirement plan.
What's more striking is that OpenAI halted development itself. We used to worry about bad actors using AI. Now the worry is AI being too capable on its own — Astra crossed the "critical" line in agentic coding, meaning it can write code and run attacks autonomously. The threat model flipped: it's no longer "someone wields AI," it's "AI is the weapon."
Lean proofs are hard currency in math — no bluffing, just formal verification. AI's math isn't hallucination; it's submitting homework that checks out. The old guard mocking AI for "just memorizing the training set" needs a new angle to be anxious from.
Source: The Hacker News、The Guardian、DataCamp