Cognition ships SWE-2 coding model: 1 point behind Fable 5.1 at 64% less cost

AxSaucedo · x · 2026-09-19

Cognition released SWE-2, its most advanced coding model, scoring 50.0% on FrontierCode 1.1 Main — within one point of Fable 5.1 while costing 64% less.
Training advance: RL scaled to the multi-trillion-parameter regime for the first time, with a new RL algorithm that trains all reasoning-effort levels in a single run, pushing the entire cost-performance frontier.
SWE-2 is post-trained from Kimi K3 (2.8T parameters, already heavily RL'd for agentic coding), adding 5–6 points on many benchmarks.
Results: 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1, beating SWE-1.7 and Grok 4.6 on both score and cost, matching GPT-5.6 Sol and Fable 5/5.1 at a fraction of the price, and nearing GPT-6 Astra at a quarter of the cost.
The post details post-training tricks like per-effort linear cost penalties tuned to local slope.

More from coding & agent