OpenAI has published a blog post claiming that Astra, a next-generation model it has not yet released, made breakthroughs on 10 longstanding mathematical problems, according to reports from MSN, digitimes and 36 Kr.
The claim did not go unanswered for long. According to MSN's report, an Anthropic employee said the company's Claude Fable model had cracked 5 of the problems — a quick counter-claim that turned a research announcement into a scoreboard between two of the field's biggest labs.
Two other details are circulating. A report from sofx.com says the 10 results were produced for roughly $2,000 in compute cost. And 36 Kr reports that OpenAI accompanied the announcement with a 249-page research paper laying out the full work.
A few caveats are worth stating plainly. Astra is unreleased, so outside researchers cannot yet run it themselves and reproduce the results. The available coverage does not specify which problems were solved, how hard they are considered within mathematics, or how much human guidance was involved. And OpenAI is the source of the claim about its own product — the 249-page paper is what independent mathematicians would need to check before treating the results as settled.
Why it matters: mathematics has long been treated as a clean test of whether AI systems can genuinely reason rather than recombine what they have read, so a credible win here — at a cost sofx.com puts in the low thousands of dollars — would suggest research-grade AI is arriving faster, and cheaper, than expected.