Cloudflare’s Auto Router Rates Each Request for Stakes, from the Most Recent Turns

Cloudflare has released Auto Router in public beta through AI Gateway, an update reported by InfoQ's Steef-Jan Wiggers. A developer sets the model to cloudflare/auto and a classifier then chooses a model for each request rather than binding a conversation to one model for its duration.
Routing happens in two stages. AI Gateway first assembles a candidate pool, dropping models that cannot handle the request format and applying credentials, billing configuration, access control and spend limits. A multi-head classification model then runs on Workers AI on Cloudflare's edge GPUs. It assigns probabilities across 14 task categories and rates the request from 1 to 5 on four dimensions: complexity, ambiguity, stakes and dependence on earlier context. A scoring matrix combines those ratings with benchmark results under the stated formula "utility = expected quality - adaptive cost penalty." Price is weighted less as difficulty rises, so harder work leans toward stronger models.
The classification reads only the newest turns, capped at 8 KiB of text. That design choice drew the sharpest practitioner pushback. Igor Pyrin, a tech lead with banking and healthcare platform experience, noted that a long agent session might establish early that a task involves a customer refund, after which the user simply says "go ahead." He suggested carrying an application-owned risk label across turns to decide whether a request may use cloudflare/auto at all, arguing that a short final instruction should not silently lower the risk classification of the underlying action. The turn that actually commits an action is often the cheapest-looking turn in the conversation.
Cloudflare published a benchmark in which its own router loses, which is the more persuasive part of the announcement. Across 97 tasks with three samples per model and 10,000 bootstrap resamples, the router trails Claude Opus 5.5 by about ten points on success rate while spending roughly a third as much. Whether that trade is good depends on what a failed request costs: as one CTO working on agentic engineering put it in a comment on Cloudflare's LinkedIn post, a cheaper model is not much of a saving if someone spends another hour fixing the result. The benchmark gives teams the numbers to make that call themselves.
Two documentation details matter for operations. Responses carry cf-aig-routed-model and cf-aig-routing-reason headers, with values such as cost_optimal_within_pool and pinned_by_turn, so a routing decision can be audited after the fact. The pool can also be narrowed per request through cf-aig-allowed-models and cf-aig-allowed-providers headers. It cannot yet be saved as reusable routing profiles, something Adam Natad asked for in the launch thread.
There is a discrepancy worth checking before trusting the figures. The default pool currently lists Claude Opus 5, Sonnet 5 and Fable 5 alongside GPT-5.6 Luna, Sol and Terra, and Grok 4.5. The two baselines in the results table, Opus 5.5 and GPT-6 Sol, sit in the opt-in list reachable only through those headers. Cloudflare says the pool may change over time, so this could be drift rather than intent, but a team trying to reproduce the benchmark should confirm what its gateway actually selects.
Sessions add cost. Without a session ID, requests in the same conversation may land on different models and cannot reuse the prompt cache. Cloudflare applies a switching penalty that grows with context length, since a newly selected model may redo reasoning tokens at output prices. In practice this suggests pinning a session when a conversation has a long shared prefix, and letting routing float across independent requests.
Critics on Reddit questioned the premise. One commenter argued auto routers end up costing more for less intelligence and that a capable model is cheaper when outcomes matter; another replied that the right unit is the task, not the request. The follow-up conceded something Cloudflare also acknowledges: classification is itself easier for a large model, and the weights go stale, especially with constantly changing models. Routing weights come from benchmark results, so new models must be added to the scoring matrix by hand.
Availability terms are limited. Auto Router is free during the beta, with no pricing announced for afterward. WebSockets are unsupported, leaving Chat Completions and the Responses API. A second router, cloudflare/auto-best, is on the roadmap; it uses the same classifier and pool but selects for highest expected quality with no cost tradeoff.
This fits a broader pattern in which gateways sell routing as infrastructure rather than as a model. The appeal is that cost control, access policies and audit headers live in one place; the risk is that a per-request decision can miss intent that only the full session reveals.
Why it matters: Teams running agents can cut spend substantially, at some cost in reliability, and the routing headers make the choice inspectable. Developers building multi-turn workflows should decide whether risk belongs in a session-level label rather than in the router's recent-turns view, since inference from this design is that a terse confirmation can understate what a request will do. Because the router is free only in beta, cost planning should assume pricing appears later.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Cloudflare has released Auto Router in public beta through AI Gateway, picking a model per request from a classifier that scores complexity, ambiguity, stakes and context dependence. Cloudflare’s own benchmark puts it ten points below Claude Opus 5.5 on success rate at a third of the cost, and practitioners have questioned how stakes survive a long agent session. By Steef-Jan Wiggers