Vercel gateway data: decision models take 26% of requests, Anthropic wins new teams 2:1

Decision models take 26% of all requests

Vercel has published its AI Gateway Production Index for November 2026, drawing on the more than 100 trillion tokens the gateway routes each month between production applications and AI labs. The headline finding: a new category of decision models, which classify, score, or pick between options instead of generating text, has grown in roughly six weeks to 26% of all gateway requests in October while accounting for just 0.53% of total spend.

According to Vercel, decision models arrived on the gateway in mid-September and there are already 15 of them from eight vendors. Nearly a third of paid teams used one in October, and about 11% used a decision model without sending a single request to a text-generating model. These models cost about a seventieth as much per request as a language model, which explains the lopsided request-to-spend ratio.

Usage skews toward the heaviest AI consumers. Just over one in five teams processing fewer than 10 million tokens a month used a decision model; at 100 million to 1 billion tokens, nearly half did; above 10 billion tokens, seven in ten did. Vercel also notes that the top 1% of teams by spend use a median of 29 different models, while the bottom tenth use one. Almost all of the request volume went to Jev from TypeSafe AI, which received nearly four times Anthropic's total requests, yet for every dollar teams spent on Jev they spent 64 on Anthropic models.

Voice adoption up 2.6x since July

About 11% of paid teams used a voice model in October, up from 4% in July, a 161% increase. Among teams that paid in both months, voice usage rose from 5% to 14%, the share of teams sending at least 100 voice requests a month grew fivefold, and six in ten of September's voice teams returned in October, up from fewer than half in July.

Vercel describes most voice pipelines as having three parts: speech-to-text, a language model, and text-to-speech. Most teams deploy only one or two of these pieces, and all-in-one models and self-assembled pipelines essentially swapped places between July and October, with assembled pipelines now the more common approach.

Anthropic and OpenAI reach parity

In December, 70 of every 100 paid teams on the gateway used an OpenAI model and 48 used an Anthropic model. By October the two were nearly level at 55 and 54. Most of Anthropic's gain came from new teams: in October it was the first-choice lab for nearly twice as many of them as OpenAI.

Among the top 1% of teams by spend, more than nine in ten already use both labs, so competition has shifted to how much work each one gets. On that measure Anthropic still leads but OpenAI has gained ground, taking nearly 15 cents of every dollar spent on the gateway, up from about 9 cents in December, while Anthropic's share halved to about 32 cents.

A six-week half-life for models

More than half of October's language-model requests went to models less than 60 days old; before September that share had never passed three in ten. Within a model family, a new version arrives every 61 days at the median, and a typical model peaks within a month of launch. Four weeks later, three in four of the teams using it at its peak have moved on, whether or not a successor has shipped.

In October, about half of teams running at least 10 million tokens a month changed their most-used model. Four in five of those moved to a more recent release, and six in ten moved to a different lab. Vercel's summary: AI models now have a half-life of about six weeks.

Also in October's data

Anthropic's Opus 5.5 was the top model by spend in its first full month, taking about one in seven dollars across all teams, and four of the six most widely used models on the gateway are Anthropic's. Image and video both had their biggest month yet, with OpenAI's GPT-Image-2.5 Sunburst leading image spend and ByteDance's Seedance 2.5 leading video spend for a third month, ahead of Google's Veo 3.1 and Kuaishou's Kling 3.0. Vercel cautions that spend is estimated at published list prices, so actual bills may differ.

Why it matters

This is a rare quantitative window into how AI is used in production rather than in demos or benchmarks. Two implications stand out. First, architecture: production AI is splitting into generation and judgment workloads, and a model category that is vastly cheaper per request changes the economics of any high-volume classification or scoring task that previously required a language model or hand-written rules. Second, vendor dynamics: with teams abandoning a peaked model within a month and switching labs six times in ten, model loyalty barely exists, so labs have to re-earn their business with every release and teams benefit from routing layers that make switching cheap. The data reflects one gateway's population of paying teams, but at 100 trillion tokens a month it is one of the larger public signals on real-world AI consumption.