Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Architecture

Jev Reached 35% of Paid Teams on Vercel's AI Gateway in Two Weeks, for 0.27% of the Spend. Its Week-Two Retention Ranked Fifth.

October 19, 2026 8 min read

On Thursday, October 15, Vercel published the October edition of its AI Gateway Production Index, a monthly report built from anonymized traffic that runs through its gateway. Vercel led with open-weight models; the coverage led with Jev, the decision model from TypeSafe AI. In its first 14 days on the gateway, Jev was used by 35.3 percent of paid teams, nearly double the 18.1 percent of Claude Sonnet 5.5 in second place. Vercel calls it the fastest adoption of any model in AI Gateway history. TechCrunch ran the story on October 16 as "Jev is Vercel's fastest-adopted AI model ever", with a fair caveat: a gateway view is at risk of undercounting labs such as OpenAI and Anthropic, which sell through their own APIs.

A disclosure before the analysis. I use Jev. My news radar asks it to rank headlines, and my posts go through a fact-checker in which Jev reads each claim against a source and answers supports, contradicts or not enough information, while code checks every number. I wrote about that pipeline on September 28. So this is a user reading a report about a tool he relies on, and I have tried to read the unflattering lines as closely as the headline.

Four numbers, not one

The reach figure is the one that traveled. The same report has three more.

Spend. Over its first two weeks, Jev accounted for 0.27 percent of gateway spend. Claude Opus 5.5, which arrived on the gateway on September 22, reached 13.1 percent of paid teams and took 11.03 percent of spend in its own first two weeks. Opus 5.5 reached a little more than a third as many teams as Jev and took about 41 times its share of the money.

Requests. Jev received more requests than any other model in September, although it launched halfway through the month. Yet it was only 1 percent of the month's tokens and 0.15 percent of its spend.

Retention. Of the teams that tried Jev in its first week, 55.6 percent used it again in week two. That ranks fifth among September launches. GPT-6 Luna led at 64.8 percent, and Sonnet 5.5 had 60.5 percent.

The composite. Vercel's full scorecard combines reach, spend and retention. Jev ranks sixth with 65.9 points. Sonnet 5.5 is first with 87.8.

So the accurate summary is longer than the headline: the most widely tried launch of the month, the most called model, almost absent from the bill, and fifth at bringing its first users back.

What reach does not measure

Read the methodology box before the headline. Launch metrics cover external paid teams during each model's first 14 days. Reach counts the teams that used the model against all paid teams active since the start of the launch month. The box sets no minimum volume for reach, so as far as it says, a team that sent one request counts the same as a team that routes every support ticket through it.

Trying Jev was also unusually cheap. The Next Web and Tech Times both reported that Vercel offered Jev for free on its gateway until September 25. I did not find that offer on Vercel's own pages, so treat it as reported. If it is right, Jev's whole first week on the gateway, which the index dates from September 17, cost nothing to try. That helps reach more than it helps retention. The index also estimates spend at published list prices, so 0.27 percent is what those calls were worth at list price, not necessarily what anyone paid.

And the gateway is one door among several. TypeSafe sells Jev directly, and Vercel's own integration guide lists Cloudflare and OpenRouter as other routes. A team that moved to a direct key after its first week looks like churn in this data.

Why the tiny share is the point

Jev's price explains most of the gap. TypeSafe charges $0.042 per million input tokens, and output tokens are free. Vercel's explanation in the report is simpler: a Jev request "is faster, cheaper and uses a fraction of the tokens."

The rest of the same report tells you what Jev is not doing. Gateway spend in September was about 13 times December's. Spend on frontier models grew tenfold over the same period. Opus 5.5 took more than one gateway dollar in ten in its first two weeks. Nothing in this data says teams are replacing their large models with a judge. My reading is that they are adding one beside them: the expensive model writes, the cheap one decides.

That matches my own setup. Jev's pass over one of my articles costs one to three cents, depending on how many sources it reads. The judge is a rounding error. The writer is the bill.

How Vercel wires it in

The more useful reading is not the index but Vercel's documentation, which shows how a platform company plugs a decision model into production.

Decision fallbacks. On AI Gateway you can make Jev the primary model and name a second model, for example GPT-6 Luna, to reconsider uncertain answers. For a Choice or Score question, the trigger is a confidence threshold. For a yes-or-no question, it is a probability band, such as 0.4 to 0.6. When the condition matches, the gateway reruns the whole state and every question with the fallback model, returns that model's complete result without merging the two, and still bills the first stage.

Two details matter if you store decisions. Only native decision models report confidence: a language-model fallback returns one value per question, so its Choice and Score answers carry no confidence. And the response names the model that produced the answers. I argued on September 29 that you should store the probability, never the verdict. With fallbacks, store the model too, and expect some rows to have no confidence at all.

The fallback band is also a budget line. On the gateway, GPT-6 Luna starts at $0.10 per million input tokens and $0.50 per million output tokens. An escalated decision pays for the state twice: $0.042 for Jev plus $0.10 for Luna, or $0.142 per million input tokens, 3.4 times a Jev-only call before Luna's output, assuming similar token counts. Escalate 10 percent of calls and the average input cost goes from $0.042 to $0.052 per million, up 24 percent. Escalate 40 percent and it reaches $0.082, nearly double. The width of your uncertainty band sets that number.

Tool approvals. When you turn on automatic tool approvals in eve, Vercel's open-source agent framework, Jev is the default reviewer. It reviews the tool name and arguments against clear and caution criteria. Clear calls run, caution calls pause for a person, and failed reviews also go to a human. Vercel's guide applies this to shell commands, to tell inspecting a file from changing or deleting it, and adds the sentence I would underline: "Keep the tool's permissions enforced when it executes." The judge is a filter, not a sandbox. InfoWorld reported on October 15 that researchers at the security firm HiddenLayer had disrupted Jev's outputs with prompt injection and adversarial text, and that TypeSafe had not responded to its requests for comment.

TechCrunch also reported that Vercel itself replaced an OpenAI Luna model with Jev to classify harmful commands, and The Next Web put the speedup at five to 18 times. I did not find that measurement published on Vercel's pages.

What I would check before wiring one in

Pin the version. TypeSafe's jev-latest alias points to jev-1.13.0, and jev-preview to jev-1.14.0, available since October 14. An alias moves when a release ships. Rerun your test cases before you move with it.

Read the defaults of your integration. Vercel's guide notes that TanStack AI turns a yes-or-no probability into a boolean at 0.5, next to the raw probability, while the AI SDK hands you the probability alone. If your code reads the boolean instead of the number, a 0.8 rule becomes a 0.5 rule and nobody notices.

Know where it runs. The gateway's model page lists two providers for Jev, TypeSafe AI and DigitalOcean, and by default the gateway picks providers on recent uptime and latency. If location matters to your clients, restrict it with the gateway's only option, and turn on Zero Data Retention, which Jev supports per request.

Measure your own week two. Vercel's changelog tells you to calibrate probabilities and confidence against labeled examples from your workflow. The index measured whether teams came back. Your version of that number is whether the decision still holds on last month's cases.

More than a third of the gateway's paid teams tried a model that cannot write a sentence. The number worth watching in November is how many still have it in the code path.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk