Cloudflare launches Auto Router public beta
Cloudflare has launched a public beta of Auto Router through AI Gateway, allowing users to set the model to cloudflare/auto and have a classifier select a model for each request. AI Gateway filters candidate models for request format, credentials, billing, access controls, and spending limits, while a multi-head classifier running through Workers AI scores task type, complexity, ambiguity, risk, and context dependence. Cloudflare reported that the router improved reliability by about 10 percentage points at roughly one-third the original cost, but its benchmark has not been independently validated.
Auto Router gives teams a built-in way to balance model quality and cost, while its auditing headers, per-request model restrictions, and limited 8 KiB context also expose practical constraints that adopters need to evaluate.

Cloudflare 最近通过 AI Gateway 推出了 Auto Router 公测版"。只需将模型设置为 cloudflare/auto,分类器就会为每个请求选择一个模型。公告中最有价值的内容,是一项 Cloudflare 自家路由器反而落败的基准测试。
路由分为两个阶段。AI Gateway 首先构建候选模型池,过滤掉无法处理当前请求格式的模型,并应用凭证、计费配置、访问控制和支出限制。随后,一个多头分类模型会在 Cloudflare 边缘 GPU 上通过 Workers AI 运行,为 14 类任务分配概率,并从四个维度对请求进行 1~5 级评分:复杂性、歧义性、风险程度,以及对先前上下文的依赖程度。最后,评分矩阵将这些结果与基准测试数据相结合,其公开公式为:“效用 = 预期质量 - 自适应成本惩罚”。