Cognition launches SWE-2: post-trained on Kimi K3, matches frontier models at up to 70% lower cost

silasalberti · x · 2026-09-11

Cognition officially launched SWE-2, calling it its "closest model yet to the frontier": on leading evals it scores on par with recent frontier models at up to 70% lower cost, via RL scaled to trillions of parameters that pushes the Pareto curve on both capability and cost.

Key details from the thread:

Scores 50.0% on FrontierCode 1.1 Main (Cognition's "would a maintainer merge this PR" benchmark); Fable 5.1 gets 50.9 at 64% higher cost, Grok 4.6 48.0, Sol 47.5, Astra 53.3
Leads Terminal-Bench 2.1 at 92.8
Revealed to be post-trained on Kimi K3 (SWE-1.7 was post-trained on Kimi K2.7)
First Cognition model with effort levels, all from a single RL run
SWE-2 medium beats SWE-1.7 in 53 steps per run vs 127, at 81% less cost

Original post →

Cognition launches SWE-2: post-trained on Kimi K3, matches frontier models at up to 70% lower cost

More from coding & agent