SWE-1.7: Frontier-Level Intelligence at a Much Lower Cost

Cognition has succeeded in achieving frontier-level intelligence at a much lower cost with SWE-1.7. This model is the result of broad improvements across the RL (reinforcement learning) pipeline, including infrastructure improvements, training stability, data quality improvements, and new techniques for handling long-horizon tasks.

The new model is now available directly in Devin's web, desktop, and CLI environments. This update includes not only improvements to the model's performance but also innovations in infrastructure, algorithms, and data processing methods.