Cognition SWE-2 is a coding model for engineering teams that want frontier-level software-engineering performance without paying frontier token prices. Released on September 10, 2026 inside Devin Desktop and CLI, SWE-2 is post-trained from Moonshot AI's Kimi K3 and routed through the Devin product rather than sold as a standalone API.
Core Features
- Reinforcement-learning post-training from Kimi K3 (a 2.8-trillion-parameter base) tuned specifically for code generation and task cost.
- Three effort levels - medium, high, and max - trained in a single run with a cost penalty, so you trade compute for harder tasks without switching models.
- Near-frontier scores: 50.0% on FrontierCode 1.1 Main (within one point of Anthropic's Claude Fable 5.1) and 92.8% on Terminal-Bench 2.1.
- Fewer steps to value: a median of 18 steps to the first real edit, down from 48 in the earlier SWE-1.7.
- Bundled access: SWE-2 runs inside Devin Pro, Max, and Teams with no separate token rate.
Use Cases
- Routine PR reviews, small features, and bug fixes where cost per task matters more than marathon agentic runs.
- High-volume coding shops already on the Devin platform that want cheaper model calls.
- Teams comparing completed-task cost rather than raw benchmark leaderboard position.
Pricing
SWE-2 has no standalone API price; it ships inside Devin. Devin Pro costs $20 per month, and Cognition offered SWE-2 free across Pro, Max, and Teams subscriptions through October 10, 2026. See related models in our AI engine and model directory.
Our Take
Best for budget-conscious teams doing short, well-scoped coding tasks inside Devin. The trade-off is clear on long-horizon work: on Terminal-Bench 4 it scores 27.3% versus 55.8% for Claude Fable 5.1, so overnight multi-file agentic sessions still belong to a frontier model.
