Cognition Releases SWE-2 Coding Model at Lower Cost
Cognition’s SWE-2 reaches near-frontier coding scores while claiming sharply lower cost and fewer turns on long software tasks.
Cognition has released SWE-2, a coding model designed for its Devin software-engineering agent and positioned as a lower-cost alternative to larger frontier systems.
Performance and training
Cognition reports that SWE-2 scores 50.0% on FrontierCode 1.1 Main, within one point of Claude Fable 5.1 in the company’s comparison, while costing 64% less. On DeepSWE 1.1 it scores 73.0%, and on Terminal-Bench 2.1 it reaches 92.8%.
The company says the model is post-trained from Kimi K3, a 2.8-trillion-parameter base model, using reinforcement learning at what it calls the multi-trillion-parameter scale. Its training method applies cost penalties across different reasoning-effort levels in one run, aiming to improve the entire cost-performance curve instead of optimizing only for a single maximum score.
Cognition also reports a substantial reduction in agent activity. On its 100-task FrontierCode evaluation, SWE-2 medium used 53 mean steps compared with 127 for SWE-1.7, and the company says average cost fell 81% on that comparison.
Why it matters
The release adds pressure to the idea that coding-agent progress must come from ever-larger general models. If Cognition’s results hold outside its own evaluations, better reinforcement learning and more efficient agent behavior could make sustained software work more economical.
The caveat is that nearly all headline numbers come from Cognition’s benchmarks and deployment environment. SWE-2’s real importance will depend on performance across unfamiliar repositories, failure recovery and the price users actually pay in Devin.