Cognition unveils swe-2: a cost-effective coding model outperforming fable 5.1 on frontiercode

Cognition, the developer of the Devin coding agent, has launched SWE-2, its most advanced coding model to date. SWE-2 is post-trained using reinforcement learning from Moonshot AI’s Kimi K3, a 2.8T-parameter model. It achieves a score of 50.0% on FrontierCode 1.1 Main, nearly matching Fable 5.1 at 64% lower cost. The model also features selectable reasoning-effort levels, all trained in a single RL run. SWE-2 builds on the infrastructure of SWE-1.7, which was trained on Kimi K2.7. Cognition scaled RL to the multi-trillion-parameter regime, using a base model with almost three times the parameters. The model’s RL algorithm trains all three effort levels in one run, with each level having its own cost penalty. Cognition reports that SWE-2 outperforms its K3 base on every benchmark, though it trails Fable 5.1 and GPT-6 Astra by about 30 points on Terminal-Bench 4. The model also demonstrates improved efficiency, with SWE-2 medium scoring higher than SWE-1.7 while taking 58% fewer turns and costing 81% less. Cognition highlights behavioral improvements such as stronger test coverage, resourcefulness when tools are blocked, and verification discipline. The company also details technical innovations like Pareto-informed cost penalties and length-weighted reward baselines to enhance performance and training stability.