SWE 2

Changes from previous version

Cognition's first multi-trillion-parameter RL run. Versus SWE-1.7, FrontierCode 1.1 Main rises from 42.0% to 50.0%, DeepSWE v1.1 from 37.7% to 73.0%, and Terminal-Bench 2.1 from 81.5% to 92.8%. Medium effort makes its first real edit after a median of 18 steps versus 48, with 58% fewer turns and 81% lower average cost on that harness.

Release Summary

Cognition's most advanced coding model, post-trained from Kimi K3 (2.8T total parameters) with a single RL run across reasoning-effort levels. Scores 50.0% on FrontierCode 1.1 Main (within one point of Claude Fable 5.1 at a claimed 64% lower cost), 73.0% on DeepSWE v1.1, 92.8% on Terminal-Bench 2.1, and 27.3% on Terminal-Bench 4. Available in Devin Desktop and CLI at launch, with Devin Web and Fusion rolling out. No public standalone API or weights.

Timeline