Cognition says SWE-2 Flash runs three times faster and costs one-fifth

Cognition reports mixed benchmark results: Flash beats SWE-2 on FrontierCode 2.0 but trails on Terminal‑Bench 4.0 and DeepSWE 1.1. The company attributes most gains to higher‑quality, diverse training data and a novel mid‑training stage before reinforcement learning. Weights are not publicly released; instead Cognition is creating a secure sharing program for trusted industry partners and research labs, citing security concerns after internal testing raised exploit scores from 11.3% to 34.8%.