Speechless and Everywhere on Vercel’s AI Gateway

Picture a developer late on a deploy night. Their support agent routes every incoming ticket through a large chat model, with a prompt that politely asks for one word back: “urgent” or “not urgent.” Most of the time it complies. Sometimes it returns a friendly paragraph explaining its reasoning, and the parser downstream falls over. So the developer changes one line in the gateway config and points that call at a model that cannot produce a paragraph at all. The bug goes away, because the failure mode no longer exists.

That scenario is hypothetical. The adoption numbers behind it are not. According to Vercel’s data, TypeSafe AI’s Jev became the fastest-adopted model in the history of Vercel’s AI Gateway, and it did so without writing a single sentence of natural language.

The adoption curve

The figures Vercel published are striking on their own terms:

• Within 24 hours of arriving on the gateway, Jev reached nearly 13% of Vercel’s paid developer teams.
• That launch-day rate was twice that of the GPT-5.6 family and more than six times that of Fable 5.1.
• Within two weeks, Jev was being used by 35% of paid teams.
• That reach was nearly double that of the second most popular model, Sonnet 5.5.

These are teams paying for infrastructure, which makes them a reasonable proxy for production intent rather than weekend curiosity. A third of them reached for a model whose entire output vocabulary is a probability, a choice, or a number.

What Jev actually does

Jev’s interface is narrow by design. You give it some text and a question. It gives you back one of three things: a yes/no probability, a pick from up to 255 options, or a score. That is the whole contract. There is no prose, no chain of thought to scrape, no “Sure, here’s the answer” preamble to strip out.

From an agent architecture standpoint, this is the interesting part. Most agent systems I study are not mainly generating text. They are making decisions. Should this tool be called? Which of these handlers owns this request? Is this retrieved passage relevant? Does this output pass a policy check? How confident should the planner be before acting? Each of those is a classification or scoring step dressed up as a conversation, because for the past few years conversation was the only interface the strongest models offered.

When you force a generative model into a decision role, you inherit a pile of overhead. You write output schemas. You add retry logic for malformed responses. You parse strings into enums and hope the casing matches. You pay for tokens that exist only to be thrown away. A model that returns a typed decision directly removes that entire layer of glue code. The developers on Vercel’s gateway appear to have noticed quickly.

Decisions as a first-class primitive

I read Jev’s numbers less as a story about one model and more as evidence for a shift in how teams build agents. The chat model was the general-purpose engine for a while because it was the only engine available. Now it looks more like one component among several, with specialized decision models handling the branching logic that holds an agent together.

This mirrors how mature software systems usually evolve. Early on, one flexible tool does everything badly enough to ship. Later, the hot paths get specialized components with tight interfaces. Routing, gating, and ranking are the hot paths of an agent loop. They run on every step, often several times per step, and they are where malformed output does the most damage.

The output format also matters for evaluation. A probability or a score can be logged, thresholded, compared across versions, and monitored for drift with ordinary tooling. Free-form text needs another model, or a human, to judge it. For teams trying to make agents dependable, a decision that arrives as a number is far easier to reason about than one buried in a sentence.

What the numbers do not tell us

Adoption is not the same as value delivered. The published data tells us how many paid teams used Jev, not how heavily, for which tasks, or whether those teams kept using it. A team that ran a single test call counts the same as a team routing every request through it. Vercel’s gateway is also one platform with its own user base, so these figures describe that population, not the industry as a whole.

There is also a design question I would want answered before betting an architecture on this approach. A fixed set of up to 255 options is plenty for routing and classification, but some decisions in agent systems are open-ended. Teams will still need generative models for planning and for anything a user reads. Jev does not replace those; it takes work off their plate.

Why this matters for agent builders

The practical lesson is simple. Audit your agent and count how many model calls exist only to make a choice. If the answer is “most of them,” you may be paying a generative model to do a classifier’s job and then writing code to undo its creativity.

The fastest-adopted model on Vercel’s AI Gateway has nothing to say. For a large share of production agent work, that turns out to be the feature developers wanted. The quietest component in the stack may end up being the one that makes the rest of it behave.

🕒 Published: October 18, 2026