Identify AI model overuse with User Insights

Cloudflare expanded User Insights, the AI Gateway feature it launched last month, with a model overkill view that flags conversations where the selected model appears more capable than the task requires. The update is free to AI Gateway users. Auto Router, which applies task and model-fit signals automatically, also reaches public beta alongside this release.

The overkill view covers cases such as simple formatting or summarization requests sent to a high-capability reasoning model, and shows which users, agents or applications produce the pattern. Cloudflare says it is not a leaderboard and does not recommend a replacement model; teams compare cost, latency, token usage and conversation turns before deciding what to change. The related Potential Savings view identifies requests that may be handled by a faster or less expensive model without compromising output quality.

Task analysis groups conversations into initial categories: coding, research, writing, summarization and data analysis. Turns analysis shows how much back-and-forth tasks require, so teams see the time, tokens and money spent before a task finishes rather than only the first request. A separate Auto Router beta uses conversation trajectory, task category, task complexity and model-fit signals to select among the models available to an application while accounting for cost, instead of requiring a routing rule per workload. Cloudflare notes it does not simply send every request to the cheapest model.

Classification runs in a dedicated Cloudflare Worker that processes eligible AI Gateway logs, examining user requests, assistant responses, tool calls and tool results, and returning a category plus a confidence score and dimensions such as task complexity, intent ambiguity, stakes and context dependence. Metadata uses Durable Objects; log bodies use R2. Classification is asynchronous, so it adds no latency to responses but User Insights is not real time: analysis may trail incoming traffic by roughly one day. Log body retention follows configured AI Gateway logging behavior.

Identity context comes from AI Gateway plus Cloudflare Access, which is free for teams up to 50 users and lets tools such as Claude Code, Codex and OpenCode inherit it automatically. Custom applications must send a stable user_id and session_id, as in the cf-aig-metadata header example.

Why it matters: Teams routing traffic through AI Gateway get free visibility into which workloads, users and agents drive model overuse, and beta testers can let Auto Router choose models automatically rather than writing a rule for every workload. Because the dashboard is aggregate and roughly a day behind, it suits reviewing usage patterns over time rather than live monitoring.