Cognition launches SWE-2, a 2.8T-parameter coding model
Original: Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra
Why This Matters
SWE-2 demonstrates frontier-level coding performance at significantly lower cost, advancing the efficiency benchmark for agentic coding models.
Cognition released SWE-2, its most advanced coding model, achieving 50.0% on FrontierCode 1.1 Main — within one point of Fable 5.1 — while being 64% cheaper. Post-trained from Kimi K3 (2.8T parameters), it uses a novel RL algorithm spanning all reasoning-effort levels in a single run.
Cognition has introduced SWE-2, a new coding-focused AI model post-trained from Kimi K3, a 2.8-trillion-parameter base model. SWE-2 achieves 50.0% on FrontierCode 1.1 Main, placing it within one point of Fable 5.1 (50.9%) and ahead of Grok 4.6 (48.0%) and GPT-5.6 Sol (47.5%), while costing 64% less than Fable 5.1. On DeepSWE 1.1, SWE-2 scores 73.0%, and on Terminal-Bench 2.1, it reaches 92.8%.
The key technical advance is an RL algorithm that trains all reasoning-effort levels simultaneously in a single run, using linear cost penalties tuned to the local slope of the base model's Pareto frontier. Cognition reports this is the first time RL has been scaled to the multi-trillion-parameter regime. The training adds 5–6 points across benchmarks over the base Kimi K3 model.
Efficiency gains are notable: SWE-2 medium takes 58% fewer turns and costs 81% less per task than SWE-1.7 on FrontierCode 1.1 Main. The model is now available in Devin Desktop, CLI, Devin Web, and Fusion.