Cognition rolls out SWE-2 coding model across Devin tools

Cognition has rolled out SWE-2 across Devin Desktop, CLI, Web and Fusion. The company says its coding model scored 50.0% on its FrontierCode 1.1 Main benchmark, near Fable 5.1's result while costing 64% less at that score. Cognition says SWE-2 was built on Moonshot AI's Kimi K3 and uses reinforcement learning across medium, high and max effort levels. It also says the medium setting used 58% fewer turns and cost 81% less on average than SWE-1.7 while matching or exceeding its benchmark score.