A model that can only pick from a short menu of answers has just out-adopted every model that can write a sonnet. Vercel published data on Thursday showing that Jev, from TypeSafe AI, was in use by 35% of its paid teams within two weeks of launch. That is the fastest start Vercel has recorded for any model, and nearly double the reach of the runner-up, Sonnet 5.5.
The caveat arrives early, as it usually does: this is a fast start by Vercel's measure, not a universal verdict. We will get to the limits of the data shortly.
What Jev actually is
Jev is a lightweight decision model. Instead of producing open-ended text, it chooses among a few predetermined outputs. That sounds like a restriction, and it is one, deliberately.
Narrowing the possible answers lets the model run cheaper and faster than general-purpose alternatives such as Sonnet. The obvious job is classification at volume, where you have a mountain of items and a short list of labels and you would prefer not to pay for eloquence.
TypeSafe AI put Jev out in limited release on September 15. In the weeks since, the approach has generated considerable buzz, enough to prompt competing offerings from OpenAI and Amazon, among others. Imitation is the sincerest form of product roadmap.
Vibes versus numbers
Buzz in AI is cheap and plentiful. Until now, the reporting around Jev leaned heavily on good vibes and light on hard figures. Vercel's data is the first real attempt to check whether the enthusiasm turned into usage.
The headline findings:
- Adoption speed: 35% of paid teams on Vercel were using Jev within two weeks of its launch.
- Lead over rivals: Its reach is nearly double that of the second most popular model, Sonnet 5.5.
- A record: By Vercel's data, no model has been adopted faster in its first two weeks.
Vercel is not just an observer here. It was an early adopter of Jev, using it to replace GPT-6 Luna for classifying harmful commands. A company reporting on a model's popularity while being one of its customers is not a scandal, merely a detail worth knowing.
What the data can and cannot see
The numbers come from Vercel's AI gateway, a unified API service comparable to OpenRouter. Teams route requests to many models through one door, and the gateway counts who walks through.
That view is necessarily partial. It is especially likely to undercount the big labs, such as OpenAI and Anthropic, which sell access through their own APIs and so see plenty of traffic that never passes Vercel's door. So "fastest adopted ever" really means fastest adopted among the people who use this gateway, as the original report itself puts it, at least by some measures.
Still, the pattern is informative. A model built for a boring, high-volume job found its audience quickly among developers who pay for tooling, which suggests there was a real itch.
Open source keeps climbing
The report also had a broader finding about the market. Open source models now account for more than 70% of the tokens processed through Vercel's gateway, and nearly 25% of total spend.
Those two figures tell a familiar story. Open source models do an enormous amount of the work and collect a much smaller slice of the money. Tokens are cheap when the model is cheap.
Not a zero-sum game
One could read that and conclude that open source is eating the frontier labs. The data does not quite say so.
Spending on frontier labs has grown tenfold, even though it now represents a smaller share of overall spend. The pie got much larger, so a shrinking slice can still be a far bigger meal. The report takes this as support for the idea that frontier models and open source alternatives may be complementary rather than competitive, with each handling different kinds of work.
The source does not specify the period over which the tenfold growth occurred, so treat that figure as directional.
What to take from it
Three things seem reasonably safe to say. A narrow, cheap model can win quickly when it fits a common job. Competitors noticed, given the offerings from OpenAI and Amazon. And the gateway data, while incomplete, at least moves the conversation from impressions to usage.
What remains unknown is whether Jev holds its position once the novelty fades, and whether the 35% figure looks as striking when measured outside one company's gateway. Early adoption is a sprint. Classification, being the kind of work nobody notices until it breaks, tends to be a marathon.
For now, the most hyped model of the autumn is, fittingly, the one that gives the shortest answers.