Cloudflare released its Auto Router in public beta through AI Gateway, a service that automatically sends each AI request to a model capable enough for the task. The company says the tool can cut costs by up to 30% compared with using only frontier models, based on its internal use through the OpenCode harness. The release matters because organizations adopting AI face rising token spend, and the router aims to reduce that spend without requiring users to pick models manually.
Users set their model to cloudflare/auto, and the router handles selection. In an internal benchmark of general knowledge work, cloudflare/auto achieved an 86.6% success rate across 252 of 291 trials, at a total cost of $2.10 and $0.0084 per success. Anthropic’s Claude Opus 5.5 scored 96.6% (281/291) at $5.91 total and $0.0210 per success, while OpenAI’s GPT-6 Sol scored 84.2% (245/291) at $2.64 total and $0.0108 per success. The benchmark used 97 tasks with three samples per model per task, and confidence intervals were estimated from 10,000 task-level bootstrap resamples. Cloudflare says Auto Router came in at 80% the cost of Sol and 35% the cost of Opus.
The router works in two stages. It first builds a pool of models that can serve the request, filtering by format, execution mode, credentials, billing, access policies, spend limits, and provider health. It then sends a compact view of the conversation to a multi-head classification model running on Workers AI and deployed on GPUs across Cloudflare’s edge network. That classifier assigns probabilities across 14 task categories and rates the request on complexity, ambiguity, stakes, and dependence on earlier context, each on a one-to-five scale. A scoring matrix combines those signals with model benchmark results to estimate fit, and the router selects the model with the highest utility, defined as expected quality minus an adaptive cost penalty.
Cloudflare says the design makes routing decisions legible, because users can inspect a task’s predicted category and complexity. Adding a new model does not require retraining; the company only adds benchmark-derived weights to the scoring matrix. The router also accounts for cache costs in long agentic sessions. Within a turn, it rarely switches models because the cache is hot. Across turns, it applies a switching penalty that grows with the number of tokens already in context, pricing a model with a live cache at its cheaper cache-read rate and every other candidate at the full cost of rewriting the context. Cloudflare notes that most models cannot read another model’s reasoning tokens, so a switch that drops reasoning tokens may force the new model to redo that work at output prices. In the future, the company wants the router to prefer staying within the same model family when it switches.
Cloudflare positions AI Gateway as a control plane for organizations deploying AI internally. Because every request from users, agents, and tools flows through it, the gateway can do more than observe and enforce budgets, spend limits, and identity-aware analytics. The company says budgets and rules still rely on individuals to make cost-conscious choices request by request, and the next step is for the gateway to make intelligent decisions on a user’s behalf. Cloudflare also plans to release other routers, including cloudflare/auto-best, which uses the same classification and model pool but selects the highest expected quality without applying the cost tradeoff.
The public beta is available now through AI Gateway, but several details remain unproven. Cloudflare’s cost savings and benchmark results come from its internal usage and its own general knowledge work benchmark, not from independent testing. The company says Auto Router does best across a wide range of knowledge-work tasks, like those in a large organization spanning technical and non-technical teams, and that its internal results for coding tasks are comparable with frontier models. Whether those results hold for other organizations, workloads, or model pools is not yet established. The release is also only the beginning of what Cloudflare says the Auto Router can learn from its position in the inference path.