OpenAI said GPT-6 Astra lowered the small prime-gap bound to 186, beating Julia Stadlmann’s 240 result and Axiom Math’s interim 212 within days of their announcements.
The record matters because pushing the bound down from 246 brings mathematicians closer to the twin prime conjecture; Stadlmann’s late-August paper was the first advance in more than a decade.
Axiom built directly on Stadlmann’s methods after she posted online, while OpenAI said it had also been exploring the problem alongside work on large prime gaps and released the 186 result with Astra on Sept. 3.
Terence Tao, Andrew Granville and others said the episode intensified fears that AI firms are chasing benchmarks over understanding, bypassing academic norms and making human work obsolete before its ideas are absorbed.
OpenAI says it does not plan to keep pushing the prime-gap result, but mathematicians are now scrutinizing its proof strategy and debating how AI will reshape access, credit and the culture of mathematical discovery.