Cognition, the company behind coding agent Devin, has introduced a new programming model called SWE-2. The model is built by continuing reinforcement learning on Moonshot AI’s 2.8 trillion-parameter Kimi K3, and Cognition said the extra training lifted results across several coding benchmarks by about 5 to 6 percentage points. On Cognition’s in-house FrontierCode 1.1 Main benchmark, SWE-2 scored 50.0%, ahead of GPT-5.6 Sol at 47.5% and Grok 4.6 at 48.0%. It trailed Fable 5.1 by 0.9 percentage points, with Fable posting 50.9%, while running at 64% lower cost. GPT-6 Astra scored 53.3%, leaving SWE-2 3.3 percentage points behind, with Cognition putting SWE-2’s cost at about one-quarter of Astra’s. The company also said SWE-2 is more efficient than its predecessor: at medium reasoning intensity, task-completion interaction rounds fell 58% and average cost dropped 81%. Results were weaker on the harder Terminal-Bench 4, where SWE-2 scored 27.3%, well below Astra’s 57.9% and Fable 5.1’s 55.8%. SWE-2 is now live in Devin Desktop and CLI, and is starting to roll out to Devin Web and Fusion.