According to CoinMeta, the parent company of Cognition, Devin, has released a new programming model, SWE-2. This model is trained using reinforcement learning based on the parameters of Kimi K3 from "The Dark Side of the Moon" version 2.8t. After training, it achieved a score of 50.0% on Cognition's Frontiercode 1.1, surpassing the 47.5% of GPT-5.6 and the 48.0% of Grok. It was only 0.9 percentage points short of the 50.9% achieved by SWE-2 on Fable 5.1, but with a 64% lower cost. Compared to GPT-6 and Astra, SWE-2 scored 3.3 percentage points less, yet its cost was about one-quarter of the latter two. At a medium level of reasoning intensity, SWE-2 reduced the number of interaction rounds required to complete tasks by 58%, and the average cost decreased by 81%. On Terminal-bench 4, SWE-2 scored 27.3%, which is lower than the 57.9% of Astra and the 55.8% of Fable 5.1. SWE-2 has already been launched on Devin Desktop and CLI, and is being rolled out to Devin Web and Fusion.