Summary
On September 10, 2026, Cognition released SWE-2, its most capable coding model, post-trained from Moonshot Kimi K3 using reinforcement learning that trains medium, high, and max effort levels in a single run. SWE-2 scores 50.0 percent on FrontierCode 1.1 Main, within about a point of Anthropic Fable 5.1, while Cognition says it costs roughly 64 percent less to run at that score. It is available in Devin Desktop and Devin CLI at launch.
What changed
Cognition launched SWE-2, a coding model post-trained from Kimi K3 with single-run multi-effort RL, available first in Devin Desktop and Devin CLI and rolling out to Devin Web and Fusion; no standalone API, pricing, or weights were published.
Why it matters
It signals that open-base post-training can land near frontier coding performance at a large cost discount, pressuring premium per-token coding models. Bundling SWE-2 into Devin tightens Cognition control of the coding-agent stack from model to execution surface.
Evidence excerpt
SWE-2 is post-trained from Kimi K3 and scores 50.0 percent on FrontierCode 1.1 Main, within a single point of Fable 5.1 while costing 64 percent less to run at that score.