transmission 06 · model wing, hall ii
October 29, 2025. A frontier-size model — hundreds of billions of parameters — served at up to 950 tokens per second on Cerebras wafer-scale silicon. SWE-1.5’s claim was not that it was the smartest model in the room; it was that the smartest model in the room had always been allowed to be slow, and that stopped here.
| 950 tok/s | 6× | 13× |
| peak serve rate | faster than haiku 4.5 | faster than sonnet 4.5 |
The stack was co-designed end-to-end — model, inference, and agent harness treated as one system. Serving came via a partnership with Cerebras, whose wafer-scale hardware runs the model at rates conventional GPU clusters don’t reach. The pitch to users was simple: you no longer choose between an AI that thinks fast and one that thinks well.
Before launch, the model beta-tested undercover. Windsurf users noticed a mysterious entry in the model picker labeled Falcon Alpha — an unannounced checkpoint collecting real-world telemetry. The annals record this as the moment the lab learned to dress its research in a trench coat.
SWE-1.5 didn’t arrive alone. Two weeks earlier — October 16, 2025 — Cognition released SWE-grep, an agentic model trained purely for rapid context engineering: the retrieval specialist that finds the right code before the flagship reasons about it. The pattern of pairing a big thinker with a fast fetcher persists in the product to this day — see the invisible models exhibit.
Developers shouldn’t have to choose between an AI that thinks fast and one that thinks well. — Cognition, October 29, 2025
Next hall: same silicon, smoother manners — SWE-1.6.
— 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.