Models
SWE-2
Cognition's most advanced coding model, released September 10, 2026. Post-trained with RL from Kimi K3, it scores 50.0% on FrontierCode 1.1 Main, within a point of Claude Fable 5.1 at 64% lower cost, and ships in medium, high and max effort levels.
Highlights
-
Frontier scores, 64% cheaper
50.0% on FrontierCode 1.1 Main, 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1, ahead of Grok 4.6 and GPT-5.6 Sol on score and cost and within a few points of GPT-6 Astra at a quarter of the price.
-
Starts editing sooner
Medium effort makes its first real edit after a median of 18 steps against 48 for SWE-1.7, with 58% fewer turns and 81% lower cost per task.
-
Three effort levels from one RL run
Medium, high and max are trained together with a cost penalty tuned to the Pareto frontier. Medium gets moving on simple tasks; high and max plan, explore and verify more.
-
Built on Kimi K3
Post-trained from the 2.8T-parameter Kimi K3, the first time Cognition scaled RL to multi-trillion parameters. The RL adds 5 to 6 points on many benchmarks.
Availability
Available in Devin Desktop and Devin CLI, rolling out to Devin Web and Fusion. There is no standalone API and no open weights, so you pick swe inside a Devin product rather than calling it from your own harness.