Cognition announces SWE-2 coding model, near-parity with Fable 5.1 at 64% lower cost
Cognition
Cognition introduced SWE-2 (Sep 10), a frontier coding model post-trained from the 2.8T-parameter Kimi K3 base and the first where the company scaled RL to the multi-trillion-parameter regime, training all reasoning-effort levels in a single run via a cost-penalty reward. It scores 50.0% on FrontierCode 1.1 Main (within one point of Fable 5.1 at 64% lower cost), 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1, and is available in Devin Desktop and CLI with Devin Web and Fusion rollout underway.
Why it matters
A second-tier lab claiming near-frontier agentic-coding performance at roughly a quarter of the cost pressures pricing across the coding-model market and shows open-weight bases like Kimi K3 as viable RL substrates.
Importance: 3/5
Notable coding-model release claiming frontier parity at lower cost