Jev reached 35 percent of Vercel teams for 0.27 percent of spend

Vercel says one AI model has been used by more of its paid teams than any launch in its gateway's history and still took less than half of one percent of their spend. The model is a decision model, which does a different job than the ones beside it, and both numbers come from the same two weeks.

In the October AI Gateway Production Index, published 15 October 2026 by Amelia Charles and Harpreet Arora, Vercel reports that Jev from TypeSafe AI reached 35.3 percent of its paid teams in its first 14 days. The second-fastest launch Vercel has recorded, Anthropic's Sonnet 5.5, reached 18.1 percent in the same window. Over Jev's first two weeks it accounted for 0.27 percent of all gateway spend.

What Jev is, in Vercel's words

Vercel describes Jev as a system-one model that picks between a few typed answers rather than writing text, so each call uses far fewer tokens than a language model answer. The post also reports 55.6 percent week-to-week retention on Jev among teams that tried it, which is Vercel's measure of how many came back the next week.

What the rest of the top five looks like

Vercel's first-14-days ranking by paid-team reach gives Jev at 35.3 percent, Sonnet 5.5 at 18.1 percent, GPT-6 Luna at 14.9 percent, Gemini 3.8 Flash at 13.2 percent and Opus 5.5 at 13.1 percent. On spend share from a September launch, Opus 5.5 is the largest, at 11.03 percent of September spend; GPT-6 Astra follows at 4.86 percent and DeepSeek-V4.1-Flash at 4.19 percent. Vercel notes that 89 percent of Opus 5.5 spend came from existing Opus 5 customers.

Open-weight share of the gateway

Vercel reports that open-weight models accounted for 71.5 percent of gateway tokens and 23 percent of gateway spend in September. Vercel says open-weight spend share has quadrupled since June, when it was 5 percent. DeepSeek-V4.1-Flash took 58 percent of all gateway tokens for 6.6 percent of spend, Vercel's own numbers for the cheapest-per-token model in the index.

What switched

Vercel says 49 percent of its tracked teams changed their primary model in September and 33 percent changed their primary lab, a record. Anthropic's gateway spend share fell from 63.7 percent to 43.9 percent, the first time it has been below half in the index.

What the numbers do not say

Vercel states its spending figures use published list prices and that actual bills may differ, and that its gateway does not see teams that call vendor APIs directly, so Anthropic and OpenAI shares are under-counted by an unknown amount. The 14-day launch window counts paid teams that touched a model, and the switching cut includes only teams with at least 10 million paid tokens in both months. The post gives no latency figure for Jev. A team deciding whether to pilot a decision model should measure that themselves.

The read for a team shipping an agent today

A decision model at Jev's price sits beside a language model in the same agent. One call writes, a cheaper call decides. That is why the reach number can be large and the spend number small over the same two weeks. The forecast line for a budget becomes two numbers for one agent: a cost for the writer and a cost for the decider. Retention of 55.6 percent on Jev in Vercel's data is high enough that teams who tried the pattern kept using it the following week.

Source

• AI Gateway Production Index, October 2026, Vercel
• Jev is Vercel's fastest-adopted AI model ever, TechCrunch