Cognition's SWE-2 chases cost per task, not the top score

Cognition released SWE-2, the coding model behind its Devin agent, and the pitch is cost per solved task rather than a new peak score. The model is built on Kimi K3, the 2.8-trillion-parameter open-weight model from Moonshot AI, which is itself worth noting: a US coding startup is now shipping a product on a Chinese open base. On Terminal-Bench 2.1 it scores 92.8%, up from 81.5% for the previous SWE-1.7, and on the harder FrontierCode 1.1 it reaches 50%, matching Anthropic's Fable 5.1 at what Cognition says is 64% lower cost.

The training work is where the detail is. Cognition used what it calls Pareto-optimized reinforcement learning, adding a cost penalty to the reward so the model improves at every effort level at once rather than trading cheap-and-dumb against slow-and-smart. The behavioral result they highlight is restraint: SWE-2 makes its first real code edit after a median of 18 steps, against 48 for SWE-1.7, which means less thrashing before it understands the task. They also describe length-weighted baselines to steady training and speculative decoding for throughput. The write-up is here.

Why it matters

If you pay for coding agents by the token, the number to watch is not the top benchmark but the cost to clear a task, and a model that matches Fable 5.1 at a fraction of the price changes the math on running agents at scale. It is also a live test of whether open Chinese base models are now good enough to build commercial products on.