Cognition SWE-2 achieves 50% on FrontierCode 1.1 at 64% lower cost

SWE-2 uses a multi-trillion-parameter RL recipe to train all reasoning-effort levels in a single run. Post-trained from the 2.8T-parameter Kimi K3, the model matches Fable 5.1 performance while reducing costs by 64% and nearing GPT-6 Astra capabilities.