The number
Vercel’s October gateway report, published on 15 October, says Jev reached 35.3% of its paid teams in its first two weeks after launching on 17 September, and calls that the fastest adoption of any model in AI Gateway history.
The next four entries in the same table are Claude Sonnet 5.5 at 18.1%, GPT-6 Luna at 14.9%, Gemini 3.8 Flash at 13.2% and Claude Opus 5.5 at 13.1%. Jev’s two-week reach is close to double the model behind it. Retention was 55.6%: a little over half the teams that tried it in week one came back in week two.
The spend figure is the other half of the story, and the report is not internally consistent about it. Its summary says Jev took 0.27% of gateway spend over those two weeks; the body text says 0.15% and the scorecard says 0.3%. The report does not reconcile the three. What all three agree on is the order of magnitude: a model on a third of paying teams is taking well under one per cent of the money.
What Jev is
Jev comes from TypeSafe AI and is a decision model rather than a text generator. Its outputs are constrained to a small set of predetermined choices — a label, a score, a yes or no — which is why it is cheap and fast enough to put in front of large volumes of data, and why it cannot hold a conversation.
Vercel is a customer as well as the scorekeeper here. It replaced GPT-6 Luna with Jev in the classifier that decides whether a command is safe to run.
That combination explains the shape of the numbers. A classifier runs on every request rather than on every conversation, so it reaches a lot of teams; and because each call is short and constrained, it costs almost nothing. Reach and revenue come apart.
The caveat on the data
These are gateway figures, not industry figures. Vercel’s AI Gateway is a routing service, and developers who call OpenAI or Anthropic directly do not appear in it, so the report undercounts the labs that sell their own API. TechCrunch, which covered the release, makes the same point.
The rest of the report is worth reading for a different reason. Open-weight models took 71.5% of tokens through the gateway in September and 23% of spend, up from about one dollar in twenty in June. Spend on frontier models has grown tenfold since December, and spend on open-weight models more than seventyfold. Average price per token fell 47%.
What to watch
Whether the spend share moves. A model that reaches a third of teams and earns a fraction of a per cent is either a very cheap commodity or the thin end of something — and the report’s own retention column, where GPT-6 Luna leads at 64.8%, is the better early indicator of which.
