Cloudflare flags when a model is overkill for the job
Cloudflare's User Insights now shows when a model is more capable than the task needs, and which users or agents are driving the spend.
Cloudflare has added a model-overkill view to User Insights, the analytics layer that sits on top of its AI Gateway, so teams can see when a conversation was handled by a model far more capable than the work required.
The company's argument is that tokens and request counts only ever tell half a story. A rising bill could mean developers are tackling harder problems, agents are making too many follow-up calls to finish a job, or a small group of users and automated agents are consuming a disproportionate share of the traffic. Numbers alone cannot separate those cases, so User Insights now tries to describe the task behind each request.
What the overkill view actually shows
The new view flags conversations where the selected model appears oversized for the task, such as formatting or summarisation requests sent to a high-capability reasoning model. It then attributes that pattern to specific users, agents or applications, so an administrator can investigate rather than guess.
Cloudflare is explicit about what the feature is not. It is not a leaderboard, and it does not automatically recommend a replacement model; it is meant to prompt questions about whether a model is being used because it fits, because it is the default, or because an agent was configured to reach for the same model at every step. The capability is available at no charge to AI Gateway users.
Our opinion
Model hygiene is the least glamorous problem in enterprise AI and the one most likely to quietly eat a budget. Routing every task through a frontier reasoning model is the software equivalent of sending a courier to post a birthday card, and most teams cannot see that they are doing it until the invoice arrives. The decision to describe intent rather than score models is the right call, because a leaderboard would simply push everyone towards the same handful of cheap names and call it optimisation. The honest test of this feature is whether it changes routing defaults, not whether its dashboard looks tidy.