Cloudflare released Auto Router for AI Gateway in public beta on September 30, 2026. AI Gateway sits between an application and its model providers; setting the model to cloudflare/auto lets it choose a model per request. The router is free during beta, while model inference remains billable. Source

Account and billing prerequisites

You need a Cloudflare account, a gateway and authentication with an API token carrying AI Gateway Run permission. That permission is account-wide, so do not treat the token as restricted to a single gateway. Source

Use your own provider credentials through BYOK, or prepay for Unified Billing credits. Unified Billing needs authentication and sufficient credits. Its endpoints prioritize a request-supplied provider key, then a stored default-alias BYOK key, then Cloudflare-managed credentials. Check Require provider credentials if requests must use your own keys. Source

Compatibility and usage limits

Current Auto Router documentation lists Chat Completions and Responses formats as supported, with WebSockets unsupported. The launch post still describes full Responses support as future work. Our editorial recommendation is to verify the exact Responses features your integration needs. Request settings can restrict eligible models and providers; cf-aig-session-id supports cache affinity across conversation turns. Source Source

Unified Billing requests using Cloudflare-managed credentials are limited to 200 per 60 seconds per gateway, returning 429 when exceeded. This particular limit excludes BYOK requests. Accounts whose first gateway was created on or after September 24, 2026 follow Workers Logs limits and pricing, so check log-storage costs separately. Source

What the vendor’s savings figures show

Cloudflare reports up to 30% savings from internal OpenCode use. A separate knowledge-work evaluation ran 97 tasks three times: Auto Router succeeded in 252 of 291 trials for $2.10; Opus 5.5 succeeded in 281 for $5.91. These are distinct vendor evaluations. Source

Assessing savings and quality for your service also requires results from its actual tasks. The second evaluation has different success counts, making cost per accepted result worth comparing alongside cost per call. OYOPICK reviewed official documentation and vendor-reported results; we did not independently benchmark Auto Router.

Hypothetical calculation using published prices

Cloudflare’s model catalog, checked October 1, 2026, lists Claude Opus 5 at $5 per million input tokens and $25 per million output tokens; Sonnet 5 is $2 and $10 respectively. Tokens are the units of text billed by the model. The following request volumes and routing mix are assumptions. Source Source

Assume 10,000 requests, each with 2,000 uncached input tokens and 500 billable output tokens. Token counts are identical for both models. Exclude retries, tools, additional reasoning tokens, cache reads and writes, taxes and logging costs.

  • All Opus 5: 20 million input × $5/million + 5 million output × $25/million = $225
  • 40% assigned to Sonnet 5: 8 million input × $2/million + 2 million output × $10/million = $36
  • Remaining 60% assigned to Opus 5: 12 million input × $5/million + 3 million output × $25/million = $135
  • Hypothetical mixed-model inference total: $36 + $135 = $171

Unified Billing adds 5% when credits are purchased. If both scenarios buy the necessary credits afresh and consume them fully, the comparison becomes $225 × 1.05 = $236.25 versus $171 × 1.05 = $179.55: a difference of $56.70, or 24%. Actual purchase increments, leftover credits and additional charges are separate considerations. Source

The 40/60 split is an invented scenario, not observed Auto Router behavior or a reproduction of Cloudflare’s 30% claim. Real spending depends on chosen models, token counts, caching and retries.

Our suggested adoption criteria

Start with a bounded comparison against the existing fixed-model baseline on representative tasks. Measure task success, cost per accepted result, slow-request latency, retries and cache behavior, and keep a route back to the existing model. Response headers including cf-aig-routed-model and cf-aig-routing-reason expose routing information. Source

For sensitive workloads, review eligible providers and data-handling terms first. The launch post puts zero-data-retention model filtering and reasoning-level selection on the roadmap. Provider retention policies and AI Gateway logging require separate checks. Source Source