Cloudflare has released Auto Router in public beta for AI Gateway, allowing applications to request `cloudflare/auto` instead of naming one model. A classifier then selects a model for each request while honoring format compatibility, provider credentials, access controls and spending limits.
The system assigns probabilities across 14 task categories and scores complexity, ambiguity, stakes and dependence on earlier context. It combines those signals with model benchmarks and a cost penalty. Response headers identify both the selected model and the routing reason, while developers can restrict the permitted models and providers for an individual request.
The central limitation is context. Classification uses only the newest conversation turns and caps that view at 8 KiB. In a long agent session, an early instruction may establish that an action involves money or sensitive data while the final message simply says to proceed. Applications handling consequential work should therefore carry their own persistent risk label instead of assuming the router will reconstruct the stakes.
Cloudflare’s test found a lower success rate than a top-end model but at roughly one-third of the cost, making the tradeoff visible rather than eliminating it. The beta does not support WebSockets, and Cloudflare has not announced post-beta pricing. Teams should also monitor the changing default model pool and use session IDs when prompt-cache reuse matters.