Cloudflare released Auto Router in public beta through AI Gateway. Developers point the model field at cloudflare/auto and a classifier decides which model serves each request.
Routing happens in two stages. Gateway first narrows to models that can handle the request format and the customer’s credentials, billing setup, access controls and spend limits. A multi-head classifier then runs on Workers AI at the edge.
That classifier assigns probabilities across 14 task categories and rates each request from one to five on complexity, ambiguity, stakes, and dependence on earlier turns.
The cap that draws criticism
Requests are classified from the newest turns only, capped at 8 KiB of text. Scoring then combines task strength with a cost penalty that shrinks as difficulty rises.
Practitioners pushed back on that window. Igor Pyrin, a tech lead who has worked on banking and healthcare platforms, asked how the router keeps an early disclosure such as a customer refund from being dropped once a user simply says go ahead.
His suggestion is to carry an application-owned risk label across the session instead of inferring stakes from the tail of the conversation.
The number Cloudflare published
Cloudflare ran the router against its own benchmarks across 97 tasks with three samples per model and 10,000 resamples. It reports about ten points of reliability for roughly a third of the spend, a tradeoff the company leaves to the buyer to take or refuse.
