Vercel finds decision models take 26% of gateway requests but just 0.53% of spending

Six weeks after arriving on Vercel’s AI Gateway, decision models were handling one in four requests. Vercel’s November 9 Production Index shows how quickly narrow, cheap judgments have found a place in production software. The October figures also put Anthropic almost level with OpenAI in paid-team reach, while attracting nearly twice as many new teams.

From September arrival to October volume

Decision models reached the gateway in mid-September. Rather than write an answer, they classify an input, assign a score or choose between options. Vercel now lists 15 such models from eight vendors, giving teams a narrower alternative to text-generating models.

The cost difference is large. Vercel puts a decision-model request at about one-seventieth the cost of a language-model request. Lower token prices and very short answers help explain why a category handling 26% of requests accounted for only 0.53% of estimated spending.

Nearly a third of paid teams used a decision model in October. About 11% used one without sending any requests to a text-generating model. Adoption was highest among heavy users: seven in ten teams processing more than 10 billion tokens that month used decision models.

Almost all decision-model request volume went to TypeSafe AI’s Jev. It received nearly four times as many requests as all Anthropic models combined. Yet for every dollar teams spent on Jev, they spent $64 on Anthropic—a sharp distinction between activity and commercial weight.

26%Share of requests

Decision models handled roughly a quarter of gateway requests.

0.53%Share of estimated spending

Their spending share remained below 1%, using published list prices.

Anthropic gains teams; OpenAI gains spending share

The provider picture has changed since December, when 70% of paid gateway teams used OpenAI and 48% used Anthropic. By October, those figures were 55% and 54%. These are overlapping groups, not competing slices of a single market: teams can use both.

New arrivals drove most of Anthropic’s gain. In October, nearly twice as many new teams picked Anthropic as their first-choice lab as picked OpenAI. Vercel defines that choice by which lab received the most language-model requests during a team’s first month—not an exclusive commitment.

Among the top 1% of teams by spending, more than nine in ten used both providers. Anthropic still led gateway spending, but its share fell to about 32% from roughly twice that in December. OpenAI’s rose from about 9% to nearly 15%.

More voice, less time with any one model

Voice also spread between July and October, rising from about 4% to 11% of paid teams. Among teams that paid in both months, adoption increased from 5% to 14%, showing that existing customers were adding voice alongside the arrival of new users.

Teams increasingly assembled separate listening, language and speaking models rather than using an all-in-one voice model. In July, about a quarter of voice teams used all-in-one models and one in six assembled the pieces. By October, those proportions had swapped.

Model choices were changing quickly, too. More than half of October’s language-model requests went to versions less than 60 days old. Among teams processing at least 10 million tokens in both September and October, about half changed their most-used model. Of those switchers, four in five moved to a newer release and six in ten changed labs.

Older versions did not disappear. Seven in ten teams processing at least 10 million tokens still used a model at least ten months old, though it handled less than 3% of the typical team’s tokens. Continued use and primary-model loyalty were different things.

The index measures anonymized, aggregate AI Gateway traffic through October 2026, not the entire AI market. Its measurement notes define paid teams as those making at least one paid request that month. Spending uses published list prices, so actual bills may differ. Model age starts when a version first appears on the gateway, rather than necessarily its original release date.