Cognition Trains SWE-2 Using Kimi K3, Performance Nears Astra at Just 1/4 the Cost

Beating AI Express: Cognition, the parent company of coding agent Devin, has launched a new programming model, SWE-2. Built directly on Moonshot’s 2.8T-parameter Kimi K3 and further trained via reinforcement learning, SWE-2 delivers a 5-6 percentage point improvement across multiple programming benchmarks. On Cognition’s own FrontierCode 1.1 Main benchmark, SWE-2 scores 50.0%, outperforming GPT-5.6 Sol (47.5%) and Grok 4.6 (48.0%). It lags Fable 5.1 (50.9%) by just 0.9 percentage points, yet is 64% cheaper. GPT-6 Astra scores 53.3%; SWE-2 trails by 3.3 percentage points while costing roughly a quarter of Astra’s price. SWE-2 also brings major efficiency gains over its predecessor: under medium inference intensity, it cuts interaction rounds for task completion by 58% and reduces average costs by 81%. However, on the more challenging Terminal-Bench 4, SWE-2 only scores 27.3%, notably lower than Astra’s 57.9% and Fable 5.1’s 55.8%, indicating it has not yet fully matched leading models. SWE-2 is now live on Devin Desktop and CLI, and is being rolled out to Devin Web and Fusion.