Cognition factored RSA-260 with 18 concurrent Devin sessions

Cognition factored RSA-260 on September 9, splitting it into two 130-digit primes. The cost sheet carries the real stakes for anyone building with agent fleets. About 4,900 GPU-days and roughly $400K at market rates, over three weeks of wall clock. An average of 3 concurrent Devin sessions, peaking at 18, steered by one researcher, Eric Lu.

The previous record, RSA-250, had stood since February 2020. Cognition scavenged the GPU time from fragmented idle nodes on its NVL72 LLM training racks.

Devin, Cognition's coding agent, wrote a GPU-optimized General Number Field Sieve implementation for the run, per Cognition's writeup. Lu wrote more than 82,000 words across 3,328 messages steering it over three weeks. That works out to about 158 messages a day, one every four minutes across an eight-hour day. He wasn't checking in periodically. He was in the loop continuously.

Lu's own account is specific about the division of labor. He says the benchmarking framework Devin built "was not otherwise going to self-assemble." The research strategy, what to measure and where to point the cluster, was his call. Devin ran the measurement infrastructure and cluster orchestration end to end.

There's a real limit on how far this generalizes. Number field sieve work splits into independent chunks with a machine-checkable answer. Most engineering doesn't have either property. Nobody should read this as proof that 18 concurrent agents will behave the same way on a codebase with tangled ownership and no tests. But if you're running parallel agent sessions at a ratio well above 3-to-1 per human and not seeing quality drop, ask why. Either your tasks are more independent than Cognition's, or you're not looking closely enough to see the damage.