Claude and Devin push the public factoring record 34 bits in 16 days

About $82 a GPU-day is what the September 3 run implies: roughly 4,900 GPU-days for an estimated compute cost of $400,000 ^[5][2]. The prize that once sat on RSA-896 was $75,000, and it lapsed when the RSA Factoring Challenge closed in 2007 ^[7], so the earlier of these two records spent something like five times a retired prize on a number nobody was paying for ^[4].

Two straight lines run out of these records, and they disagree by a factor of about 69 ^[8]. Both lines are extrapolations, and the dollar cost of each additional bit decides which one holds. Date the curve from RSA-250, the 829-bit number factored by purely classical methods in 2020 ^[9], and the public record has gained 67 bits in six years, about 11 a year ^[5]; 1,152 bits still separate 896 from the 2048-bit keys in production ^[13], which at that rate is 103 years ^[6]. Date it from September 2026 instead, 34 bits in 16 days ^[1][2], and the same distance closes in about 18 months ^[7].

cryptobriefing.com writes that each additional bit roughly doubles the difficulty, and that 2048-bit numbers therefore remain orders of magnitude harder than anything broken so far ^[8]. Apply that. Thirty-four bits is a factor of about 17.2 billion ^[3], and 17.2 billion times $400,000 is $6.9 quadrillion ^[9]. Nobody spent $6.9 quadrillion between September 3 and September 19. So the same heuristic being used to put 2048 safely out of reach does not describe the general number field sieve across that range.

I think the repricing here is narrow. Claude, on cryptobriefing.com's account, helped Weis navigate the parameter space of a computation that a small team of number theorists would have needed months to configure manually ^[15]. If that is what happened, the binding input on factoring records has been specialist labour, and a record every 16 days says that labour got cheaper, not that the sieve got faster. The counter-thesis is that the compute got cheaper too, in which case each record is a procurement decision and the curve everyone is drawing is a spending curve. The publisher also says the RSA-896 result does not change the fundamental security analysis for properly configured modern systems ^[12].