Het verhaal
On 10 Sep 2026 Cognition introduced SWE-2, its most advanced coding model, post-trained from Kimi K3 (2.8T) with multi-trillion-parameter RL that trains all reasoning-effort levels in one run via Pareto-informed cost penalties. Headline scores: FrontierCode 1.1 Main 50.0% (Fable 5.1 50.9%, GPT‑6 Astra 53.3%) at a claimed ~64% lower cost than Fable 5.1; DeepSWE 1.1 73.0%; Terminal-Bench 2.1 92.8% (table high); Terminal-Bench 4 still weak at 27.3% vs Fable 5.1’s 55.8%. Versus SWE-1.7, SWE-2 medium scores higher on FrontierCode while using ~58% fewer turns and ~81% less average cost, with first real edit after a median 18 steps (was 48). Available today in Devin Desktop and CLI; Web and Fusion rollout underway.