Cognition's SWE-2 lands within a point of Fable 5.1 on FrontierCode at 64 percent lower cost

Cognition introduced SWE-2, which it calls its most advanced coding model, on September 10. The company positions the release as a push on the cost-performance frontier rather than a raw capability record. SWE-2 scores 50.0 percent on FrontierCode 1.1 Main, within one point of Claude Fable 5.1's 50.9 percent, while costing 64 percent less. It trails GPT-6 Astra's 53.3 percent by a few points at roughly a quarter of the cost. On DeepSWE 1.1 it scores 73.0 percent against 74.1 percent for Astra, and it reaches 92.8 percent on Terminal-Bench 2.1. Cognition says the model beats its predecessor SWE-1.7 and Grok 4.6 on both score and cost, and matches GPT-5.6 Sol and Fable 5 and 5.1 at a fraction of their price. The model is post-trained from Kimi K3, a 2.8-trillion-parameter model that had already undergone extensive reinforcement learning for agentic coding. Cognition says this is the first time it has scaled RL to the multi-trillion-parameter regime, building on the SWE-1.7 training infrastructure. The key addition is an RL algorithm that trains all reasoning-effort levels in a single run, which the company says shifts the whole cost-performance curve rather than one point on it. Its RL still adds five to six points over K3 on many benchmarks. The release comes days after Cognition raised over $2 billion at a $48 billion valuation.