OpenAI announced on August 1 that its new Astra model had solved 10 longstanding problems in mathematics and theoretical computer science. The reaction from the mathematical community was swift, and it wasn’t applause.
Prominent researchers accused the company of downplaying prior work, failing to properly attribute existing contributions, and overstating the novelty of its AI-generated results.
The claims and the pushback
According to OpenAI, the Astra model tackled problems spanning high-dimensional sphere packing bounds, the establishment of non-sofic groups, coding theory advances, and arithmetic circuit complexity. The company framed these as areas where there had been “no progress on the main result for at least a decade.”
Steven Miller, a mathematician at Yeshiva University, and Francesco Fournier-Facio from the University of Cambridge were among the researchers who pushed back hard. Their argument: the Astra results built on prior work that OpenAI either ignored or underrepresented. Andreas Thom and Miller himself had published relevant research as far back as 2016, work that directly informed the territory Astra was now claiming to have conquered.
The total compute cost for all 10 results was reportedly around $2,000 in token usage.
By August 6, Scientific American had published a report characterizing OpenAI’s handling of the announcements as “research misconduct” and plagiarism. OpenAI subsequently revised some of its statements.
A pattern, not an incident
This isn’t OpenAI’s first brush with controversy over mathematical claims. Back in October 2025, former VP Kevin Weil posted that GPT-5 had solved problems related to Paul Erdős’s conjectures. That post was quietly deleted.
Then came May 20, 2026, when OpenAI claimed its model had disproved a conjecture dating back to 1946. That announcement received endorsements from respected mathematicians including Noga Alon and Thomas Bloom, lending it genuine credibility.
Why this matters beyond academia
The $2,000 compute cost figure is a perfect example of how context shapes meaning. In isolation, it’s a remarkable data point suggesting that frontier mathematical research might soon become radically cheaper. In context, it raises a different question: what exactly did that $2,000 buy? If Astra synthesized decades of human work into a neat package without crediting the humans, the compute cost is less a measure of AI capability and more a measure of how cheaply existing knowledge can be repackaged.