Vercel reports rapid uptake of Jev decision model
Vercel says Jev, the decision model created by Typesafe AI, was used by 35% of paid teams on its AI Gateway within two weeks of launch. The company says that made Jev its fastest-adopted model in the first two weeks of availability, with nearly twice the reach of the next most popular model, Sonnet 5.5.
Jev entered limited release on September 15. It is designed to choose among a small set of predetermined outputs rather than produce open-ended text, positioning it for classification tasks across large volumes of data.
A narrow model design for classification
Restricting the output set is intended to make Jev cheaper and faster to run than alternatives such as Sonnet. That trade-off is central to its use case: organisations can apply a decision model where the job is to assign a known label or route an item, rather than generate a broad natural-language response.
Vercel was itself an early adopter. It used Jev to replace GPT-6 Luna for classifying harmful commands, offering a concrete example of the type of bounded workflow the model targets. The adoption data suggests that interest in this approach translated into usage among teams accessing models through Vercel’s service.
What Vercel’s data does and does not show
The figures come from Vercel’s AI Gateway, a unified API service comparable with OpenRouter. They therefore describe activity inside that gateway, not the entire model market. Vercel notes that the view can undercount large providers, including OpenAI and Anthropic, when customers use those companies’ direct APIs.
The results also sit alongside Vercel’s broader production AI direction: Vercel’s open-protocol production AI strategy frames the company’s emphasis on open protocols for production use, while its gateway data shows open-source models processing more than 70% of tokens. Those models account for nearly 25% of total spend through the gateway.
Open-source growth has not eliminated spending on frontier labs. Vercel reports that spending on those labs rose ten-fold even as their share of overall spending declined. The pattern supports the possibility that frontier and open-source models are being used for different needs rather than simply replacing one another.
Practical implication for teams
For businesses, Jev’s uptake is a reason to evaluate constrained-output models where the required outcome is already known, especially in high-volume classification. Any production choice should compare task accuracy, response time, cost and the coverage limits of the available usage data before replacing an existing model.