Cognition has launched SWE-2, a high-performance coding model post-trained from the 2.8T-parameter Kimi K3. By scaling reinforcement learning (RL) to the multi-trillion-parameter regime, SWE-2 optimizes the Pareto frontier of capability and cost. A key technical innovation is a novel RL algorithm that utilizes Pareto-informed cost penalties, applying a linear penalty per effort level tuned to the local slope of the base model's frontier. This allows the model to be trained across all reasoning-effort levels in a single run. To stabilize training, Cognition employed a length-weighted reward baseline to reduce gradient variance. Furthermore, the team improved RL rollout serving via DSpark speculative decoding and online draft-model training, while utilizing NVFP4/FP8 kernels and quantization-aware training to reduce memory usage. On the FrontierCode 1.1 Main benchmark, SWE-2 achieved a 50.0% solve rate, performing within one point of Fable 5.1 while remaining 64% cheaper. On DeepSWE 1.1, it reached 73.0%, matching GPT-5.6 Sol at a fraction of the cost. These advancements result in a model that demonstrates superior engineering judgment, characterized by more focused codebase exploration and improved verification discipline compared to its predecessor, SWE-1.7.