Cognition releases SWE-2 with adjustable effort for coding tasks

Cognition introduced SWE-2 on September 10, a coding model with adjustable reasoning effort. It is available in Devin Desktop and CLI, with rollout planned for Devin Web and Fusion. Cognition built it by further training Kimi K3 for software engineering.

On Cognition's FrontierCode 1.1 Main benchmark, SWE-2 scored 50.0%, close to Claude Fable 5.1's 50.9%, at an estimated 64% lower cost. On DeepSWE 1.1, it scored 73.0% against Fable 5.1's 67.4%. The comparison reverses on Terminal-Bench 4: SWE-2 scored 27.3%, while Fable 5.1 reached 55.8%. These figures come from Cognition's evaluations of those specific tasks.

The company also measured how many steps its agents took. SWE-2 at medium effort averaged 53 turns per FrontierCode run, down from 127 for SWE-1.7. Cognition attributes much of that reduction to more focused code exploration, with fewer repeated reads before the first edit. Higher effort levels spend more steps on planning and verification.