Cognition announced SWE-2, saying it matches recent frontier models on leading benchmarks at up to 70% lower cost, achieved by scaling reinforcement learning to trillions of parameters.