User Insights groups conversations into task categories such as coding, research, writing, summarization and data analysis, and its overkill view flags simple work like formatting sent to high-capability models. It does not automatically recommend a replacement model.
Classification runs asynchronously in a dedicated Worker after AI Gateway logs each request, so it adds no latency, but the dashboard can trail live traffic by about a day. Cloudflare positions it for spotting patterns over time, not live monitoring.
The classifier reads the whole conversation trajectory, including tool calls and results, and returns a category, a confidence score, and ratings for task complexity, intent ambiguity, stakes and context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → dependence.
Turns analysis shows how much back-and-forth each task needs. Cloudflare's guidance is that a simple task that keeps taking several turns points to a prompt, model or workflow problem, since the first request is only part of the cost.
Custom apps must send a stable user_id and session_id for analysis. Putting AI Gateway behind Cloudflare Access, free for up to 50 users, lets Claude Code, Codex and OpenCode inherit identity automatically.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Why it matters
Helps teams find where expensive models run simple tasks and which users or agents drive that spend, so model choices can be changed on evidence.
