Vercel says Jev set a gateway adoption record, reaching 35% of paid teams

Jev’s pitch was a smaller, cheaper job for AI: choose among fixed answers rather than write a response. Vercel’s October 15 usage report now shows that approach finding a wide audience. The model reached 35.3% of paid teams in its first two weeks on AI Gateway, a platform record—but captured just 0.27% of spending.

A record for reach, not repeat use

TypeSafe AI’s Jev reached nearly twice as many paid teams as Claude Sonnet 5.5, which placed second at 18.1%. Vercel calls Jev the fastest-adopted model in the gateway’s history. The comparison measures each model’s first 14 days, rather than giving an earlier release a longer window to attract users.

The model also received more requests than any other model during September, despite arriving halfway through the month. That adds a separate measure of activity: Jev did not merely appear across a large number of teams; it also led the gateway in calls sent to a model.

Repeat use tells a less dominant story. Of the teams that tried Jev in week one, 55.6% used it again in week two, the fifth-highest return rate. GPT-6 Luna led that measure at 64.8%. Jev’s record therefore concerns how broadly it caught on, not the strongest evidence that initial users came back.

35.3%Paid-team reach

The highest first-14-day reach in the gateway’s history.

0.27%Gateway spending share

Jev’s share of paid gateway spending during its first two weeks.

Fixed answers keep the bill small

Jev entered limited release on September 15; Vercel lists its gateway launch as September 17. It is a decision model, designed to select from a few predetermined outputs. The narrow task suits classification—sorting inputs into defined categories—rather than the open-ended writing associated with a general-purpose language model.

Vercel says a Jev request is faster, cheaper and uses a fraction of the tokens of a typical language-model request. Tokens are the units of data processed by a model. That design helps explain why a large number of calls need not produce a large share of either token traffic or spending.

Across September as a whole, Jev accounted for 1% of gateway tokens and 0.15% of spending. Those monthly figures use a different window from the launch scorecard. For comparison, Claude Opus 5.5 took 11.03% of gateway spending in its own first two weeks, the largest share among September launches.

One concrete use was inside Vercel itself. TechCrunch reports that the company replaced GPT-6 Luna with Jev for classifying harmful commands. That is a bounded judgment, matching the model’s limited-output design rather than asking it to take over every function of a broader language model.

The gateway is not the whole market

Vercel’s index uses anonymized, aggregate traffic routed through AI Gateway, a service offering access to multiple model providers. It does not measure all AI use. TechCrunch notes that this view may undercount labs such as OpenAI and Anthropic because customers can also reach their models through those companies’ own interfaces.

The launch metrics cover external paid teams. Reach uses the paid-team population active from the start of the launch month through the data cutoff; spending share uses total paid gateway spending during each model’s launch window. Vercel estimates that spending at published list prices, so it is not a record of customers’ actual bills.