Vercel's October AI Gateway Production Index, published Thursday and reported by TechCrunch on October 16, contains a result that should recalibrate what builders think a winning model launch looks like. Jev, a small decision model from TypeSafe AI released in limited form in September, reached 35.3 percent of paid teams on Vercel's AI Gateway in its first two weeks, nearly double the 18.1 percent reach of Anthropic's Sonnet 5.5 in second place. Vercel calls it the fastest adoption of any model in the gateway's history, and Jev logged the most requests of any model in September despite launching mid-month, with 55.6 percent of first-week teams returning in week two.
Here is the twist: that record-breaking reach produced almost no revenue. Jev accounted for just 0.27 percent of gateway spend in its first two weeks. That is because Jev is not a general-purpose chatbot competing on reasoning benchmarks. It is a decision model, built to choose quickly between a few predetermined outputs. Constraining the answer space makes each call faster, cheaper, and far lighter on tokens than asking a frontier model to generate a response, which is exactly what high-volume classification work needs. Vercel liked it enough to use it itself, replacing GPT-6 Luna for classifying harmful commands.
The same report shows the wider market splitting in two directions at once. Open-weight models now process 71.5 percent of gateway tokens and take nearly one in four dollars of spend, a share that has more than quadrupled since June. September gateway spend was roughly 13 times December's level while the average price per token fell 47 percent, and almost one in three teams changed their most-used lab in a single month, an index record. Yet spending on frontier models still grew tenfold over the same period. As Vercel itself puts it, frontier and open models are looking complementary rather than purely competitive: expensive models take the hard reasoning, cheap models take the volume.
There are honest caveats. A gateway is a biased sample: it undercounts labs whose customers mostly arrive through their own APIs, and reach is not the same as importance. A model that touches a third of teams for a fraction of a percent of spend is infrastructure, not a flagship. But that is precisely the lesson for practitioners. The model is becoming a replaceable part inside a larger system, and the teams processing the most tokens on Vercel already juggle dozens of models. Routing each job to the smallest model that can do it reliably is now a core engineering skill, not an optimization afterthought.
If you build with AI, the practical move is to audit your own traffic the way Vercel audits its gateway. Separate the requests that genuinely need a frontier model from the classification, routing, filtering, and extraction jobs that a narrow decision model can handle at a sliver of the cost. Jev's numbers suggest the market has already figured this out: the fastest-adopted model of the year did not win by being the smartest in the room. It won by doing one small job, very cheaply, very fast, millions of times.
