Can Automatic Model Routing Cut Your AI Bill? Test Cloudflare’s Beta on Real Tasks
News / AI costs
The router is free during beta, but model usage and failed attempts still belong in the cost comparison.
Source information checked on October 3, 2026 at 07:56 (UTC+8). This brief uses official documentation and stated editorial calculations; the site has not tested the product or verified an individual account’s eligibility or bill.
Cloudflare released AI Gateway’s Auto Router in public beta on September 30. Requests using cloudflare/auto can be assigned to eligible models according to task fit and cost. Cloudflare says the router is free during beta and reports savings in its own internal use; those results do not establish a saving for another business. Official announcement
For a small team paying for repeated AI requests, the useful experiment is straightforward: run the same tasks with your current model and with routing. Include routine categorization, difficult exceptions and longer conversations.
Current documentation supports Chat Completions and Responses formats; WebSockets are not yet supported. With session affinity configured—or a supported client that configures it automatically—the same model can serve a turn to reuse the prompt cache. Without that setting, requests may go to different models. Official documentation
Measure total provider charges, retries and corrections per accepted result. A lower-priced response is less useful if someone must repair it repeatedly. Check which providers are eligible before sending client material.
Our automation cost guide gives you a place to compare the whole task, rather than treating the router’s beta price as the entire bill.
Limits and unverified details
Savings are the vendor’s internal report, not a result measured by this site. Router-free beta does not include model usage. The beta end date, later router price, regional eligibility, providers and data-processing conditions are unverified.