Cloudflare AI Gateway Auto Router public beta (`cloudflare/auto`)

Cloudflare put Auto Router into public beta through AI Gateway (blog 2026-09-30): set model to cloudflare/auto and the gateway picks a model capable enough for the request so callers need not choose manually (blog, docs).

This is a Desk Bot tools/agents briefing. Story is per-request routing on the same Gateway control plane—not a new budgets/limits feature.

How routing works

Gateway builds an eligible pool (request format/inputs, credentials, billing, spend limits; unhealthy providers dropped). A Workers AI multi-head classifier scores 14 task categories plus four 1–5 dimensions (complexity, ambiguity, stakes, context dependence); a scoring matrix ranks quality vs price (utility = expected quality − adaptive cost penalty). Top model is tried first, with fallback if the provider cannot serve (blog, docs).

Enable path

Unified OpenAI-compatible Chat Completions (and Responses API): send model: "cloudflare/auto" to the gateway compat endpoint. Response headers include cf-aig-routed-model, cf-aig-routing-reason, and cf-aig-routing-decision-id. WebSockets API is not yet supported (docs).

For agent/coding turns, cf-aig-session-id pins the model for a turn (user message + tool follow-ups) to keep prompt cache hot; optional cf-aig-turn-id / cf-aig-no-session-affinity. Narrow the pool with cf-aig-allowed-models / cf-aig-allowed-providers. Clients such as OpenCode can send session IDs automatically; docs ship an opencode.json + @cloudflare/aig-opencode-plugin path (docs).

Default pool (docs; may change): includes Anthropic Claude Fable/Opus/Sonnet 5, OpenAI GPT-5.6 Luna/Sol/Terra, xAI Grok 4.5; additional models (Workers AI, Alibaba, Fireworks, etc.) only when allow-listed. Do not freeze specific SKUs as permanent defaults.

Pricing (beta)

Blog: “The Auto Router is free while in beta.” Upstream inference is still billed per provider/gateway usage as usual—do not imply all token spend is free (blog). No invented GA date or post-beta Auto Router fee.

Soft: CF-internal results (attribute)

All figures below are Cloudflare-internal / self-reported—not independent validation (blog):

• Early internal OpenCode harness: up to ~30% cost savings vs frontier-only (OpenAI Sol / Anthropic Claude Opus).
• Internal general knowledge-work table (291 trials): cloudflare/auto 86.6% success / $2.10 total vs Opus 5.5 96.6% / $5.91 and GPT-6 Sol 84.2% / $2.64.

Roadmap teases (not shipping)

Near-term plans called out on the blog include expanding models, ZDR filtering, provider capacity, reasoning-level selection, fuller Responses/WebSockets support, and a future cloudflare/auto-best (quality without cost tradeoff)—roadmap only, not GA claims (blog).

Who should care

Teams already on AI Gateway who want coding-agent / org traffic off a permanent Opus default should start at the Auto Router blog and docs—keep beta + free-router / billed-inference, soft-attribute CF-internal benches, and treat the default pool as mutable.