AI applications are beginning to split routine decisions from generative work. In its November AI Gateway Production Index, published November 9, Vercel reports that specialist “decision models” handled 26% of requests passing through its gateway in October, while accounting for just 0.53% of estimated spending.

That contrast is the headline, but it needs context. Vercel's sample covers traffic routed through its own AI Gateway, not the entire AI market. The company defines decision models as services that classify, score or select among options rather than produce lengthy prose. The report identifies the category as a recent addition to the gateway and notes that its observed traffic is concentrated in one model. One system, TypeSafe AI's Jev, supplied most of the category's traffic.

Smaller jobs, different economics

Traditional chat models can make a routing choice or assign a category, but they are designed to generate sequences of text, often with a correspondingly larger token bill. A decision model can answer a narrow classification or selection question with a brief output. Vercel describes such requests as cheaper within its observed traffic, although price is not the same as accuracy or suitability for a workload.

The numbers suggest a practical architectural shift. An application might use a low-cost decision model to direct a support ticket, choose a workflow or classify a document, then invoke a more capable generative model only when a task needs explanation or reasoning. That is an example of how teams could use the category; the report does not show that every observed request followed such a sequence.

The report says adoption rose among paying AI Gateway teams and was especially visible among its largest customers. Large customers also route traffic across more distinct models, so the figures should not be read as proof that decision models outperform general-purpose AI at complex tasks.

Voice grows while models turn over faster

The same report offers two other snapshots of production AI. Vercel says use of voice models increased between July and October. Meanwhile, more than half of language-model requests in October went to versions less than 60 days old. For a typical model, Vercel says three in four teams using it at peak were no longer using that version four weeks later.

Those figures reinforce how quickly developers are revisiting model choices. They do not necessarily mean that most companies have replaced their entire AI stack, nor that the newest model is always the best model for a specific workload. Older versions can remain in service for stability, testing or cost reasons.

Vercel also reports that Anthropic attracted nearly twice as many first-choice new paid teams as OpenAI in October. That metric measures gateway usage, not total model revenue or user adoption across the industry.

What buyers should measure

For organizations implementing agents, the useful question is less “Which model leads the leaderboard?” than “What must this individual step accomplish, and how can it be checked?” A decision model might reduce latency and inference expense for high-volume routing. It should still be measured for false classifications, consistency, recoverability and whether downstream actions have appropriate human approval.

The primary index and Vercel's methodological overview provide the underlying definitions. Independent reporting by TechRadar has separately highlighted the unusually fast turnover. The spending figures are estimates based on gateway data, and they should not be mistaken for all customers' actual invoices or industry-wide market share. The data is a valuable observation of one gateway—not a census of AI production worldwide.