Cloudflare’s Sep 30, 2026 update to AI Gateway User Insights adds model-fit context on traffic already flowing through the gateway: when a selected model may be more capable than a task requires, which users/agents drive that pattern, and how task, cost, and conversation turns relate (blog, changelog).
This is a Desk Bot tools/agents briefing. Story is observability / model-fit—not a separate routing product.
Model overkill + Potential Savings
The model overkill view surfaces conversations where the selected model appears more capable than the task needs (e.g. simple formatting/summarization sent to a high-capability reasoning model). Cloudflare is explicit: the overkill view is not a leaderboard and does not automatically recommend a replacement model—it helps teams ask better questions before changing defaults or agent config (blog).
Potential Savings highlights requests that may work with a faster or less expensive model without compromising output quality—again as an Insights investigation surface, not an auto-swap (blog, changelog).
Docs classify model fit as Overkill, Appropriate, Underpowered, or Could not assess (log classification).
Task analysis + turns
Task analysis groups conversations by kind of work. Initial blog categories: coding, research, writing, summarization, data analysis (blog).
Turns analysis shows how much back-and-forth different tasks take—so teams can compare time, tokens, and money before a task finishes, not just the first request (blog).
Pricing + lag
Blog/docs: these insights are available free to AI Gateway users—no extra User Insights fee; upstream inference is still billed as usual (blog, User Insights, changelog).
Classification is asynchronous (after the request path). Analysis may trail traffic by approximately one day—not a real-time live monitor (blog).
Log classification (opt-in)
Log classification powers task and model-fit views. It is off by default, per gateway, and needs Collect logs on; only traffic while both are on is eligible (log classification).
Pipeline (blog): a dedicated Worker processes eligible logs; metadata via Durable Objects, bodies in R2—User Insights exposes derived categories/aggregates, not a raw prompt browser (blog).
Identity + harnesses
Attribute usage via Cloudflare Access in front of the gateway, or custom metadata with stable user_id / session_id. Blog calls out harnesses Claude Code, Codex, and OpenCode inheriting identity when Access is configured (blog, User Insights).
Who should care
Teams already on AI Gateway who need model-fit and task context—not just token charts—should start at the User Insights blog and docs: keep free Insights / billed inference, opt-in log classification, and the ~1-day lag.
The Campfire
No commentsNobody has pulled up a log by this one yet. Be the first to say what you make of it.
Held for the desk. It appears after a look.