2026-09-11 / Signal #6
Cognition launched SWE-2, a coding model that matches frontier performance at a fraction of the cost
“A $20-a-month intern just got as good as the $200 models.”
The Story
Cognition released SWE-2, post-trained from the 2.8-trillion-parameter Kimi K3. It scores within a point of Claude Fable 5.1 and GPT-6 Astra on key coding benchmarks while claiming 64% lower cost. It is available inside Devin (including unlimited on the $20 plan for a month) but not as a public API.
Why It Matters
The coding-model race just became a price war. Builders can now rent near-frontier software engineering for the cost of a streaming subscription.
Evidence
Cognition blog and X (Sept. 10, 2026); independent benchmarks. Primary URL: https://cognition.com/blog/swe-2 (via reporting)
Sources
Daily scan: 2026-09-11