Jev reached 35% of Vercel paid teams in two weeks

Vercel says Jev reached nearly 13% of paid teams within 24 hours of its September 17 AI Gateway launch, and 35% within 14 days.[1][2] Vercel says that first-day reach was twice the GPT-5.6 family’s and more than six times Fable 5.1’s; TechCrunch reported that Jev’s reach was nearly double Sonnet 5.5’s.[2][3] Vercel also says every other recent model launch remained below 7% after a full day.[2] The figures describe reach among paid teams on AI Gateway, not all model users.

TypeSafe introduced Jev on September 15 as a model for structured software decision-making, two days before its AI Gateway launch.[2][1] Instead of returning free-form text, Jev produces typed choices, scores or Boolean answers with probabilities, and evaluates all questions in a request in parallel.[4] Vercel identifies classification, routing, scoring and automated verification as use cases.[4] Vercel has announced Jev 1.14 through AI Gateway; the version accepts images and includes an experimental long-context mode.[4]

TypeSafe charges $0.042 per million input tokens for Jev and does not charge for output tokens, according to The Next Web.[5] TypeSafe says Jev was up to 194 times faster and 445 times cheaper than language models in its own workflow evaluations.[2] Those are vendor-reported results from TypeSafe’s evaluations, not a general benchmark covering all tasks. Vercel says it used Jev instead of GPT-6 Luna to classify harmful commands.[3] That example shows a specific deployment, but does not establish that Jev replaces other models across broader workloads.

Vercel says Jev received more AI Gateway requests than any other model in September, despite launching halfway through the month.[1] Of teams that tried it in its first week, 55.6% returned in the second week, according to Vercel.[1] The company cautions that whether this early adoption lasts remains open.[2] Vercel also says its AI Gateway data may undercount activity at major AI labs that offer their own APIs, so the reported reach is not a complete measure of model use.[3]