Cognition Launches SWE-2, a Coding Model Built on Kimi K3 That Undercuts GPT-6 Astra on Price

Cognition released SWE-2, a new coding model trained on top of the 2.8-trillion-parameter Kimi K3 base, claiming comparable performance to Fable 5.1 at 64% lower cost and a quarter of GPT-6 Astra's price, rolling out now in Devin Desktop, CLI, Web, and Fusion.