Cognition has released SWE-2, a new coding model built upon the 2.8T-parameter Kimi K3 model. Initial testing on FrontierCode 1.1 Main shows SWE-2 achieving a score of 50.0%, within one point of Fable 5.1. This represents a significant cost reduction of 64% compared to Fable 5.1.
SWE-2 distinguishes itself with selectable reasoning effort levels, a feature implemented through a single reinforcement learning run. This approach utilizes a Pareto-informed cost penalty system, where the reward is defined as success minus a lambda times cost. The lambda value is tied to the local slope of the base model’s Pareto curve, ensuring that the reward rises only by pushing the frontier upwards.
Operational details reveal that SWE-2 is currently deployable only through Devin: Desktop and CLI. There are no open weights or a standalone API available. Cognition reports a 58% reduction in turns and an 81% reduction in cost compared to SWE-1.7 on FrontierCode 1.1 Main. The model demonstrates improved resourcefulness and verification discipline, as evidenced by a median of 18 steps to produce an edit, compared to 48 steps for SWE-1.7.
Cognition’s data strategy includes tripled RL environments, instruction-following overlays, and a flywheel mechanism leveraging earlier SWE-2 checkpoints to address false positives and negatives in verifiers. Furthermore, the model demonstrates strong performance on politically sensitive questions about China, passing 98.0% of questions across multiple languages. Source: https://www.marktechpost.com/2026/09/12/cognition-releases-swe-2-a-kimi-k3-post-trained-coding-model-that-matches-fable-5-1-on-frontiercode-at-64-lower-cost/
