---
title: "Cloudflare User Insights: when your AI model is overkill"
description: "Cloudflare's AI Gateway now flags when an AI model is more capable than the task needs. How it works, what it means for cost, and what to audit in your team."
date: 2026-11-03T00:00:00.000Z
---
**User Insights** is a view in Cloudflare's AI Gateway that shows who in your company uses AI, for which task and with which model. The news, announced on the [Cloudflare blog](https://blog.cloudflare.com/ai-model-overuse-user-insights/) on September 30, 2026, is the "model overkill" flag: conversations where the chosen model is more capable than the task required, such as a text summary sent to a large reasoning model. The feature is free for anyone already using AI Gateway.

In practice, it is the first time a major infrastructure vendor hands over, for free, a report that answers the question every AI budget owner asks at some point: are we paying premium prices for cheap tasks?

## What is AI Gateway User Insights?

User Insights is an analysis layer on top of AI Gateway logs (AI Gateway is Cloudflare's proxy that sits between your apps and providers like OpenAI, Anthropic and Google). Until now, the dashboard showed model names and request counts. Now it groups usage by task, model, turn and user.

According to Cloudflare, these are the new views:

| View | What it shows |
| --- | --- |
| Model overkill | Conversations where the model seems more capable than the task requires |
| Task analysis | Usage split by task type: coding, research, writing, summarization, data analysis |
| Turns analysis | How many exchanges each task type needs, with time, tokens and cost across the workflow |
| Potential Savings | Requests a faster or cheaper model could handle without losing quality |
| Identity | Usage by person, team and tool, including Claude Code, Codex and OpenCode |

Cloudflare is explicit about what the feature is **not**: it is not a model leaderboard and it does not swap models for you. It points to where you should look. The decision stays with your team.

## How does Cloudflare tell that a model is overkill?

Cloudflare classifies the whole conversation, not the single request. A dedicated Worker reads eligible AI Gateway logs **after** the response has been delivered, so the analysis adds no latency for the user.

The classifier looks at the conversation trajectory: requests, responses, tool calls and results. For each conversation it returns a work category, a confidence score and four assessments: task complexity, intent ambiguity, stakes and context dependence. An internal email summary scores low on complexity and low on stakes. If it went through a frontier reasoning model, it lands on the overkill list.

Two technical details matter if you have to take this to legal or security:

- **The dashboard shows categories and aggregates, not prompts.** Metadata lives in Durable Objects and log bodies in R2, stored separately, and retention follows the settings you already have in AI Gateway.
- **It is not real time.** Analysis can trail traffic by about a day. It is built to spot patterns over weeks, not to fire alerts.

## Who can see usage by person and by tool?

Per-person usage works when AI Gateway sits behind Cloudflare Access, Cloudflare's access control product, which is free for teams of up to 50 users. In that setup, usage from tools like Claude Code, Codex and OpenCode shows up tied to each person and team.

For your own applications the rule is different: the app has to send a stable `user_id` and `session_id` with every call. Without them, the dashboard knows what was asked but not who asked, nor that two calls belong to the same conversation.

Working at a multinational taught me that a report nobody can attribute to a team never changes a budget. If your app does not send identity, User Insights stays a curiosity.

## Does the Auto Router already pick the model on its own?

The Auto Router is the piece that turns diagnosis into action: it uses the same signals (trajectory, category, complexity and model fit) to send each request to the right model, weighing cost without just picking the cheapest one.

Its status, however, is unclear. Cloudflare's own post says closed beta in one passage and a public beta launch in another. Before you plan around it, check your account dashboard to see whether it shows up and under what terms.

## What should you audit after turning on User Insights?

Start with high-volume, low-complexity tasks, because that is where an oversized model costs the most. Summarization, classification, field extraction and simple translation are the usual suspects.

A plan for the first four weeks:

1. **Week 1:** turn the feature on and change nothing. Remember data arrives about a day late.
2. **Week 2:** open the Potential Savings view and list the five task categories with the most flagged requests.
3. **Week 3:** pick one and test a smaller model on a real sample, comparing its output side by side with the current model.
4. **Week 4:** if quality holds, switch the model for that task only and track the category's cost in Turns analysis.

One task at a time. Switching models everywhere at once is the fastest way to learn, through user complaints, which task actually needed the big model.

If you use AI to produce content, you can ask the same question outside Cloudflare: does the model that fixes a meta description need to be the same one that [reviews an entire technical document](/en/blog/deep-review-google-revisao-de-documentos-com-ia/)? In most SEO workflows I see, it does not.

## How does User Insights fit with the rest of Cloudflare's AI stack?

User Insights closes a run of Cloudflare launches aimed at teams running AI in production. [Clef](/en/blog/clef-cloudflare-modelos-de-decisao-agentes-de-ia/) returns typed decisions instead of text, [Kitesurf](/en/blog/kitesurf-cloudflare-navegador-para-agentes-de-ia/) is a browser built for agents, and [Cache Transcoding](/en/blog/cache-transcoding-cloudflare-zstandard-cache-cdn/) shrinks cache size. The logic is the same across all four: Cloudflare wants to be the layer AI traffic flows through, and measuring that traffic is the argument for staying on it.

There is a flip side worth keeping in mind. The more your cost analysis depends on one vendor's gateway, the more it costs to switch gateways later. For most small teams the visibility is worth it. For teams with their own observability, User Insights is a second opinion, not a replacement.

## Summary: what to know about Cloudflare User Insights

- **User Insights** is a free AI Gateway view that shows AI usage by task, model, turn and user.
- The **model overkill** flag points to conversations where the model is more capable than needed, and the **Potential Savings** view lists where a cheaper model would do.
- Classification runs after the response, adds no latency and can lag by about a day.
- The dashboard shows categories and aggregates, not prompts; retention follows your AI Gateway settings.
- Per-person usage requires AI Gateway behind Cloudflare Access (free up to 50 users) or a `user_id` and `session_id` sent by your app.
- The **Auto Router**, which picks models automatically, has a beta status that Cloudflare itself describes inconsistently.

What to do now: if your AI calls already go through AI Gateway, turn on User Insights today and come back in a week. If they do not, start with the simplest inventory, a spreadsheet with task, model and monthly volume, because the question Cloudflare automated can be asked by hand.