AI Can Now Do Original Science. The Cost of Breakthroughs Just Fell.

OpenAI's next model, called Astra, just solved ten mathematics problems that experts could not crack for at least a decade — some for much longer. This is not AI summarizing things it learned from books. AI can now produce results that did not exist before. That is a meaningful shift.

The problems came from deep mathematics: things like how to pack balls as tightly as possible in a thousand dimensions, or how to make certain types of encryption codes more efficient. Astra solved all ten for about $2,000 in compute time — roughly what a small business pays for a month of software subscriptions. Every solution is published as a 249-page document on GitHub, with every proof step checked by machine using a tool called Lean 4. A computer verified every single line, so no one has to take OpenAI's word for it.

The cost is the real story. Tim Gowers, a mathematician who has won the highest prize in his field, said he would recommend one of the proofs for a top journal. That kind of expert-level quality normally means years of work from a small specialist team. $2,000 for ten research breakthroughs is a number that changes how you think about knowledge work. Industries that already depend on formal verification — chip design, cryptography, safety-critical software — can now run that checking automatically.

The point is not that AI will replace researchers. It is that AI can now help find answers that no human had yet found. If your business depends on R&D — in any field where being right and being able to prove it matters — the price of discovery just fell dramatically. Start thinking now about where AI fits into your research process, not after a competitor already has.