Summary:

Vercel’s November index separates AI traffic volume from estimated spending, offering a useful lesson for teams deciding which tasks need generation and which need a constrained decision.

Vercel’s November 9 AI Gateway Production Index reports that decision models handled 26% of requests in October while representing 0.53% of estimated spending. Those are shares of different totals: heavy traffic does not necessarily mean a large bill.

What the numbers measure

The report covers anonymized, aggregated traffic through Vercel AI Gateway, not the entire AI market. Spending is calculated from published list prices, so customers’ actual bills can differ. Its definition of a paid team requires at least one paid request during the month; the report does not supply an absolute team count alongside these shares.

Vercel classifies decision models as systems that select, score or categorize rather than produce extended text. Nearly a third of paid teams used that category in October. These adoption and traffic measures do not establish whether a particular model will be accurate enough for a different workload.

A better budgeting question

For a product team, the useful question is not simply which model receives the most requests. Start with the output the application needs. Routing a support ticket into a fixed set of queues is a different job from drafting a response that explains an unfamiliar problem.

Before changing a production route, write down the allowed labels and the cases that should be escalated. Test examples near the boundaries, including ambiguous inputs. A fast answer in the wrong category can create more downstream work than it saves.

  • Measure a completed task: include retries, fallback calls and human corrections, rather than counting only successful API responses.
  • Keep a quality threshold: compare error types and their consequences alongside latency and cost.
  • Preserve a fallback: uncertain or unfamiliar cases should have an explicit destination instead of being forced into an unsuitable label.

The same discipline applies when interpreting usage charts: request counts, token volumes, spending and team adoption answer different questions. Keep their denominators visible before using a popularity figure to justify an architecture change.

Sources

Vercel: AI Gateway Production Index, November 2026, published November 9 and covering October. Featured chart: original PromptsZone visualization of the report’s category-level figures; no third-party artwork reproduced.