transmission 07 · model wing, hall iii
Spring 2026. The training runs got two orders of magnitude more compute, and Cognition spent the surplus on something unusual: manners. SWE-1.6 was post-trained to jointly optimize intelligence and ‘model UX’ — how the model feels to work with, not just what it scores.
Cognition broke its own tradition and published mid-run results. The SWE-1.6 Preview checkpoint — post-trained from the same base as SWE-1.5 — was already *+11% on SWE-Bench Pro* over its predecessor while running at the same 950 tok/s. The stated reason for the peek: get real users on the checkpoint early and tune behavior with feedback, not vibes.
The eval table in that post is a time capsule of the frontier in early 2026: Claude Opus 4.6 and Sonnet 4.6, GPT-5.3-Codex and Codex-Spark, GLM-5, Kimi K2.5, Composer-1.5 — every model run through the same three harnesses (Claude Code, Cascade, Devin) where vendors hadn’t reported numbers.
April 2026: SWE-1.6 shipped to everyone in Windsurf with a pricing stunt — free for three months. Two serving paths:
SWE-1.6 is optimized for both intelligence and model UX — trained to make the model feel smoother to use, in addition to raw intelligence. — Cognition, April 2026
| field | entry |
|---|---|
| announced | preview early 2026 · GA april 2026 |
| base | same pretrained model as SWE-1.5 |
| compute | two orders of magnitude more RL than the 1.5 run |
| headline number | +11% SWE-Bench Pro over SWE-1.5 (preview ckpt) |
| serve paths | fireworks 200 tok/s (free) · cerebras 950 tok/s (fast) |
Next hall: the model moves inside the agent harness itself — SWE-1.7.
— END OF TRANSMISSION · CTR-(ok)VERIFIED% —
+++ COGNITION ANNALS +++ DEVIN · THE AI SOFTWARE ENGINEER +++ SWE-1 → SWE-1.5 → SWE-1.6 → SWE-1.7 → SWE-2 +++ WINDSURF FOREVER +++ COMPILED BY DOGBED, NOT WRITTEN BY HAND +++ NO JAVASCRIPT WAS HARMED IN THE MAKING OF THIS ARCHIVE +++
rendered by dogbed — a document compiler, not a CMS. one binary, one template, thirteen documents.