Jev reaches 35% of Vercel's paid teams within two weeks of launch

Vercel data published on Thursday shows that 35% of paid teams on its platform used the decision model Jev within two weeks of its September 15 limited release. Vercel calls it the fastest adoption any model has recorded in its first two weeks, and the figure nearly doubles the reach of the second most popular model, Sonnet 5.5.

The numbers come from Vercel's AI gateway, a unified API service comparable to OpenRouter. The dataset gives the industry its first hard usage figures for Jev, the lightweight decision model from Typesafe AI that has drawn outsized attention since launch and prompted competing offerings from OpenAI and Amazon.

Until now, according to the report, the buzz around Jev had little data behind it. Vercel's figures indicate the attention converted into real deployment.

What is Jev, and why did it spread so fast?

Jev arrived in limited release on September 15. Typesafe AI designed it as a lightweight tool for choosing between a few pre-determined outputs. That constraint is the model's core trade: by limiting what the model can produce, it runs cheaper and faster than alternatives such as Sonnet.

The design suits one job above all — classifying large quantities of data quickly. That is a high-volume, low-glamour workload where cost per call and latency dominate the buying decision, and it is the category where a small, specialized model can displace larger general-purpose systems.

The approach has attracted significant buzz in the weeks since release, inspiring competing offerings from OpenAI and Amazon, among others, the report notes. Vercel's data now supplies evidence that the buzz tracked actual usage rather than discussion alone.

How reliable is the measurement?

The Vercel gateway view is necessarily limited. As a unified API service, it risks undercounting the major labs — OpenAI and Anthropic among them — that sell access through their own APIs and therefore do not route traffic through third-party gateways.

The measurement still carries weight for two reasons.

• It captures paying teams, not trial users or benchmark chatter.
• It reflects a two-week window, the standard early-adoption curve that Vercel uses for comparison across models.

On both counts, Jev sets the record: no model has reached a comparable share of paid teams that quickly in Vercel's data.

Vercel itself bet on Jev early

Vercel did not merely measure Jev's adoption. The company was an early adopter itself, deploying the model to replace GPT-6 Luna for classifying harmful commands.

That detail matters. Content and command classification is a security-adjacent function where errors carry real cost, and Vercel moved it from a frontier-class model to a lightweight decision model. The switch is a concrete instance of the broader pattern the report documents: specialized small models taking over workloads that once defaulted to the largest available system.

What does the report say about open-source models?

The same publication shows continuing momentum for open-source models on Vercel's platform.

• Open-source models now account for more than 70% of the tokens processed through the Vercel gateway.
• They represent nearly 25% of total spend.

The gap between those two shares is the story. Open-source models process the majority of traffic while commanding a quarter of the money, which reflects their cost advantage per token and their fit for high-volume workloads — the same niche Jev targets.

Frontier models and open source: rivals or complements?

The report's most consequential finding may be that the market is growing faster than it is consolidating. Spending on frontier labs has grown ten-fold, even as frontier models account for a smaller share of overall spend on the gateway.

That combination — shrinking share, exploding absolute dollars — gives credence, in the report's words, to the idea that frontier models and open-source alternatives may turn out to be complementary rather than competitive. Buyers are not substituting one tier for the other. They are running more of both.

The pattern fits Jev's own trajectory. A cheap, narrow decision model can classify harmful commands at scale while frontier systems handle the complex tasks they were built for. Vercel's own migration from GPT-6 Luna to Jev for classification work is the model of that division of labor.

Why the numbers matter beyond Vercel

Adoption data of this kind remains scarce in the AI market, where vendors announce capabilities faster than independent usage figures appear. The Vercel gateway is one of the few public windows into what paying teams actually run.

For Typesafe AI, the report converts two weeks of buzz into a documented record: fastest two-week adoption Vercel has measured, 35% of paid teams, nearly double Sonnet 5.5's reach. For OpenAI and Amazon, which have already shipped competing offerings, the data quantifies the market they are chasing.

The caveat stands. The gateway undercounts labs with their own APIs, so the shares describe the gateway's population of buyers, not the whole market. But within that population, the direction is clear: cheap, constrained, specialized models are winning the highest-volume work, and the money flowing to frontier labs is growing regardless.