gpt-6IMAGE
GPT-6 Astra tops ErdosBench math benchmark
GPT-6 Astra xhigh ranked first on ErdosBench, a set of 226 open Erdős-style math problems, though it beat GPT-5.6 Sol xhigh by only 5%–10%.
Przemek Chojecki | PC
@prz_chojeckiArchived 2026.09.14Originally posted 2026.09.08
Przemek Chojecki | PC @prz_chojeckiOriginally posted 2026.09.08
GPT-6 Astra xhigh scored 1st on our ErdosBench, a benchmark of 226 open math problems inspired by Erdos problems. Interestingly, it's not that big of a leap compared to GPT-5.6 Sol xhigh. Astra had a couple more problems settled than Sol with stronger scientific writing, less overclaiming (even underclaiming sometimes). A solid 5%-10% gain on various math-research skills tested. Also great to see, that ErdosBench is far from saturated. Bring on GPT-7.
x.com/prz_chojecki/status/2097278340029649202Archived 2026.09.14 15:53