Vercel data shows decision model Jev reached 35% of paid teams within two weeks

Vercel has published usage figures showing that Jev, a decision model from TypeSafe AI, has posted the fastest two-week adoption of any AI model the company has measured. According to TechCrunch, which reported the data on Thursday, 35% of Vercel's paying teams were already routing work to Jev within two weeks of its September 15 limited release. That is nearly twice the reach of Sonnet 5.5, the second most popular model on the platform over the same window.

A model built to pick, not to write

Jev is deliberately narrow. TechCrunch describes it as a lightweight model designed to choose among a small set of preset outputs rather than generate open-ended text. Constraining the answer space lets it run faster and at lower cost than larger models such as Sonnet, which makes it a natural fit for classifying large volumes of data at high throughput. That positioning has drawn heavy attention since launch and has already spurred rival offerings from OpenAI and Amazon. Until Vercel's numbers landed, however, most of the evidence for Jev's momentum was anecdotal.

Reading the fine print on the data

The figures come from Vercel's AI gateway, an aggregation layer that lets developers call many providers' models through a single API, in the same vein as OpenRouter. That vantage point has limits: as TechCrunch notes, it is likely to undercount big labs such as OpenAI and Anthropic, whose customers often connect through first-party APIs instead. Even with that caveat, the data points to genuine production usage rather than just hype. Vercel itself is one of the early adopters, having swapped out GPT-6 Luna in favor of Jev for the job of flagging harmful commands.

Open source models keep climbing

The same report charts steady gains for open source models, which now process more than 70% of the tokens flowing through Vercel's gateway and account for close to a quarter of total spend. That shift has not come at the frontier labs' expense in absolute terms: TechCrunch reports that spending on frontier models grew roughly tenfold over the period, even as their slice of overall spend shrank. The report reads this as support for the idea that frontier and open source models will end up complementing each other as the overall market expands.

Why it matters

Independent adoption data for AI models is rare, and Vercel's gateway offers one of the few quantified views of what developers actually send traffic to. Jev's climb suggests that narrowly scoped models with a clear speed and price advantage can capture real workloads far faster than general-purpose assistants, especially for commodity tasks like classification where a frontier model is overkill. For platform and tooling teams, it is a signal that routing between many specialised models, rather than defaulting to a single flagship, is becoming standard practice. The open source share figures point the same direction: the aggregator layer is increasingly where model choice, and model spending, actually gets decided.