Cloudflare launches Auto Router for AI Gateway
Cloudflare has released Auto Router in public beta, a feature that sends each AI Gateway request to a model capable enough for the task, so organizations can cut token spend without users choosing models.
Cloudflare announced Auto Router in public beta on Thursday. Developers turn it on by setting the model to cloudflare/auto in AI Gateway. The company says that internally, in its OpenCode harness, the router has cut costs by up to 30% compared with using only frontier models such as OpenAI's Sol and Anthropic's Claude Opus.
Cloudflare also tested the router on its internal knowledge-work benchmark: 97 tasks covering email, calendars, Slack, files, travel and finance, with three runs per task:
| Model | Success rate | Total cost | Cost per success | |---|---|---|---| | cloudflare/auto | 86.6% | $2.10 | $0.0084 | | Claude Opus 5.5 | 96.6% | $5.91 | $0.0210 | | GPT-6 Sol | 84.2% | $2.64 | $0.0108 |
The router cost about 80% as much as Sol and 35% as much as Opus. However, Opus finished about 10 percentage points more tasks. The router's confidence interval overlaps with Sol's.
Each request goes through two steps. First, AI Gateway filters the pool of models by request format, credentials, billing, access policies, spend limits and provider health. Then a multi-head classifier running on Workers AI reads the most recent conversation turns. It assigns probabilities across 14 task categories and scores the request from one to five on complexity, ambiguity, stakes and dependence on earlier context. Cloudflare says the router aims to minimize the predicted cost of completing the whole task, not just the price per token, because cheaper models can use more tokens to finish the same job.
The source text cuts off partway through describing how candidate models are scored. Pricing and the list of supported models were not included.