Cloudflare has expanded AI Gateway User Insights to show when an organization may be assigning a model that is more capable and expensive than a task requires. The dashboard links that pattern to users, agents and applications, then lets teams compare cost, latency, token use and conversation length.
Task analysis initially groups conversations into categories including coding, research, writing, summarization and data analysis. A separate view highlights simple work, such as formatting or short summaries, that is being sent to a large reasoning model. Cloudflare stresses that this is an investigation aid rather than an automatic verdict or model leaderboard.
The classification worker examines eligible logs, including requests, responses, tool calls and results. It assigns a task and evaluates complexity, ambiguity, stakes and dependence on prior context. Processing happens asynchronously after the gateway serves the request, so it adds no response latency but is not real time; new activity may take about a day to appear.
Derived categories and aggregate views appear in the dashboard rather than a raw prompt browser. Underlying log retention still follows each customer’s AI Gateway settings, which teams should review before enabling analysis. The feature is free for AI Gateway users and can feed Cloudflare’s separate Auto Router after teams validate a recurring pattern.