The fastest-adopted AI model ever doesn't talk

Every month, Vercel's AI Gateway routes tens of trillions of tokens between production apps and AI labs, and it publishes what it sees. The October 2026 index — data through September, published October 15 — ranks every September model launch by reach, spend, and retention. Jev, launched into the gateway September 17, topped the reach table: 35.3% of paid teams used it in its first two weeks, nearly double Claude Sonnet 5.5's 18.1% in second place.

This is the fastest adoption of any model in AI Gateway history, per Vercel's own data, and it stuck: 55.6% of the teams that tried Jev in week one came back in week two, the fifth-best retention of the month. It also fielded the most requests of any model in September despite launching halfway through the month. One caveat Vercel and TechCrunch both note: this is gateway traffic only, so it undercounts labs like OpenAI and Anthropic whose customers mostly use their own APIs.

Jev is a different kind of model. It doesn't generate conversational text. You feed it application state plus a set of typed questions, and it returns typed answers — a choice, a score, a boolean — each with an attached probability, evaluated in parallel. TypeSafe calls it a “System One model,” after fast, intuitive human thinking, and it's named after Jevons paradox: when decisions get dramatically cheaper, you make far more of them.

The pitch is simple: most of what agents actually need from AI isn't prose, it's thousands of small judgment calls — ticket routing, tool selection, risk scoring, output verification. Run those through a frontier chat model and you pay for open-ended generation you never use. TypeSafe's pricing reflects that: $0.042 per million input tokens with output free, and the company reports decisions in 70-500 milliseconds. Vercel itself was an early adopter, swapping GPT-6 Luna for Jev on classifying harmful commands.

The team has pedigree: founder Diogo Almeida is a former OpenAI researcher credited as a co-inventor of the RLHF technique behind ChatGPT. TypeSafe spent roughly two years in stealth, launched Jev in limited release on September 15, 2026, and announced a $40 million seed round led by DCVC the same day.

Here is the twist Vercel's scorecard makes plain: the model everyone tried is not the model anyone paid for. Because a Jev request is faster, cheaper, and uses a fraction of the tokens of a chat-model call, Jev accounted for just 1% of September's gateway tokens and 0.15% of spend — despite the most requests.

The model that captured the money was Anthropic's Claude Opus 5.5: 11.03% of September gateway dollars in its first two weeks, the largest spend share of any September launch — more than double GPT-6 Astra in second place. GPT-6 Luna took the retention crown at 64.8% week-one return rate. So the market is splitting: teams experiment with everything, run the cheap decisions at scale, and still write the big checks to the frontier labs.

The wider index shows the same split at scale. Open-weight models now process 71.5% of all gateway tokens but take just 23% of spend — and their share of spend has quadrupled since June. Total September gateway spend was about 13 times December's, while the average price per token fell 47%. Frontier-lab spend grew tenfold since December; open-weight spend grew more than seventyfold.

The two sides aren't substitutes. In September, 68% of teams that put more money into open-weight models did the same for closed-weight models. And Anthropic's grip on the gateway's money slipped: its share of total gateway spend fell from 63.7% to 43.9%, the first time it has dropped below half. The model is becoming a replaceable part — companies spend more on AI without becoming more dependent on any single lab.