Cognition's SWE-2 nears frontier coding models at a fraction of the cost

Cognition, the company behind the Devin coding agent, has released SWE-2, its strongest coding model yet. It scores 50.0% on the FrontierCode 1.1 Main benchmark — within a point of Fable 5.1 but 64% cheaper — and the company positions it as pushing the cost-versus-capability Pareto frontier rather than chasing raw benchmark leadership. On FrontierCode and DeepSWE 1.1, SWE-2 reportedly beats the earlier SWE-1.7 and Grok 4.6 on both score and price, matches GPT-5.6 Sol and the Fable 5 line at lower cost, and lands a few points behind GPT-6 Astra while costing a quarter as much.

The model is post-trained from Kimi K3, a 2.8-trillion-parameter model already tuned for agentic coding, marking Cognition’s first push into the multi-trillion-parameter regime for reinforcement learning. The central technical claim is a single RL run that trains every reasoning-effort level at once, using a linear per-effort cost penalty tuned to the local slope of the base model’s frontier — an approach the company argues follows from first principles because only a linear penalty behaves consistently before and after averaging cost. Supporting changes include a length-weighted reward baseline for training stability, tripled RL environments with instruction-following overlays and a self-hardening verifier flywheel, and serving optimizations using NVFP4/FP8 kernels and quantization-aware training to hold throughput despite the roughly 3x larger base.

The practical payoff is efficiency: SWE-2 medium scores higher than SWE-1.7 while using 58% fewer turns and costing 81% less, and it makes its first real code edit after a median of 18 steps versus 48 for its predecessor — a direct response to complaints that SWE-1.7 over-explored simple tasks. Cognition also reports stronger end-to-end test writing, more resourcefulness when a path is blocked, and better verification discipline (re-deriving conclusions rather than agreeing with a user’s hypothesis). SWE-2 is available now in Devin Desktop and CLI, with rollout underway on Devin Web and Fusion.