Cognition has released SWE-2, its newest coding model for Devin. It is available in Devin Desktop and the Devin CLI, with rollout to Devin Web and Fusion underway.
The model is post-trained from Moonshot AI's Kimi K3, a 2.8-trillion-parameter base. Cognition says SWE-2 is its first model trained with reinforcement learning at multi-trillion-parameter scale, using separate cost penalties for different reasoning-effort levels so the training objective rewards both successful completion and efficiency.
Cognition reports a 50.0% result on its FrontierCode 1.1 Main benchmark, compared with 50.9% for Claude Fable 5.1, and says SWE-2 costs 64% less at that comparison point. Those are vendor-reported results rather than independent benchmarks. The same Cognition table also shows a much weaker 27.3% on Terminal-Bench 4, versus 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra.
The mixed benchmark picture matters. SWE-2 looks competitive on some coding and terminal evaluations while trailing frontier models substantially on the newer Terminal-Bench 4. AiToolMap therefore treats Cognition's cost and performance claims as useful product evidence, not as independently validated proof of general parity with frontier models.
For Devin users, the material change is straightforward: Cognition now has a newer in-house model aimed at reducing the cost and number of agent steps for routine software work while preserving higher-effort modes for harder tasks.