Cloudflare has released Auto Router in public beta for AI Gateway: a classifier on Workers AI scores each request for task type and difficulty and selects a model from a configured pool, weighting price less as difficulty rises. In Cloudflare's own benchmark, the router trailed Claude Opus 5.5 on success rate by ten percentage points while costing about a third as much, and beat GPT-6 Sol on both measures.
Routing runs in two stages. AI Gateway first builds a candidate pool, filtering out models that cannot serve the request format and applying credentials, billing configuration, access control, and spend limits. A multi-head classification model then runs on Cloudflare's edge GPUs, assigning probabilities across 14 task categories and rating the request from 1 to 5 on complexity, ambiguity, stakes, and dependence on earlier context. A scoring matrix combines those ratings with benchmark results for each model, under what the post calls a simplified form of the selection logic: "utility = expected quality - adaptive cost penalty." Developers opt in by setting the model to cloudflare/auto.
The Auto Router classifies each request from the newest turns only, capped at 8 KiB of text, then scores candidate models on strength by task and expected gain on harder work, with price weighted less as difficulty rises. (Source: Cloudflare blog post)
The benchmark ran 97 agentic tasks in a simulated workspace of email, calendar, Slack, files, travel, and finance, with three samples per model and 95 percent confidence intervals from 10,000 bootstrap resamples.

The router succeeded on 252 of 291 trials, or 86.6 percent, for a total cost of $2.10. Claude Opus 5.5 succeeded on 281, or 96.6 percent, for $5.91. GPT-6 Sol succeeded on 245, or 84.2 percent, for $2.64. Per successful trial, that is $0.0084 for the router against $0.0210 for Opus and $0.0108 for Sol. Jacques B, a CTO working on agentic engineering, replied on Cloudflare's LinkedIn post that "a cheaper model isn't much of a saving if someone has to spend another hour fixing the result," which is the trade the table makes explicit.
The classification input drew the sharpest practitioner response. According to the documentation, the router classifies from a compact view that prioritizes the newest turns, capped at 8 KiB of text. Igor Pyrin, a tech lead who has worked on banking and healthcare platforms, argued that this lets a short final instruction lower the risk rating of the action it commits:
In a long agent session, the user may say "this is a customer refund" early and later just say "go ahead." I'd carry an application-owned risk label across turns and use it to decide whether the request can use cloudflare/auto at all. A short final instruction should not silently lower the risk classification of the underlying action.
The documentation gives operators two controls. The response carries cf-aig-routed-model and cf-aig-routing-reason headers, with values such as cost_optimal_within_pool and pinned_by_turn, so each choice can be audited afterward. The pool can also be narrowed per request with cf-aig-allowed-models and cf-aig-allowed-providers, though not through saved routing profiles, which Adam Natad asked for in the launch thread.
Those headers matter for reproducing the benchmark, because its two baselines are not in the default pool. The default list currently includes Claude Opus 5, Sonnet 5 and Fable 5, GPT-5.6 Luna, Sol and Terra, and Grok 4.5; Opus 5.5 and GPT-6 Sol sit on an opt-in list reachable only through the allow headers. Cloudflare says the pool "may change over time," so a team repeating the test should confirm which models its gateway can actually select.
Switching models mid-conversation has its own cost. Requests in the same conversation may land on different models and lose the prompt cache, and most models cannot read another model's reasoning tokens, so a switch can force the new model to redo that work at output prices. The router applies a switching penalty that grows with the number of tokens already in context, and keeps the same model within a turn while its cache is warm.
On Reddit, commenter reachjalil rejected the premise, arguing that auto routers "end up costing you more for less intelligence" and that a capable model is cheaper where outcomes matter; commenter yangqi answered that the unit of cost should be the task, not the request. reachjalil's follow-up raised a point the post concedes: routing weights come from benchmark results, so each new model has to be added to the scoring matrix by hand, and the weights go stale "especially with the constantly changing models."
Auto Router is free during the beta, with no pricing stated for afterward. WebSockets are not yet supported, leaving Chat Completions and the Responses API, and a cloudflare/auto-best variant that selects for highest expected quality without the cost penalty is planned but not released.