13% of Paid Teams in 24 Hours: Jev Becomes the Fastest-Adopted Model in Vercel Gateway History

Five days after an ex-OpenAI researcher quietly shipped a model that refuses to write a single word, that model has broken an adoption record. Vercel reported on September 18 that Jev, the “System One” decision model from TypeSafe AI, became the fastest-adopted model in the history of its AI Gateway: within 24 hours of listing, nearly 13% of paid teams were already calling it — more than twice the reach of any previous launch on the platform.

The numbers are stark. Jev passed every other comparison model within its first twelve hours and kept widening the lead for the rest of the day, reaching a tenth of teams in 18 hours. Every other recent model launch on the gateway remained below 7% after a full day. Measured against familiar names, Jev’s day-one share was roughly 2x the GPT-5.6 family and more than 6x that of Anthropic’s Fable 5.1.

What Jev actually is

Jev, introduced on September 15, is not a chatbot and not a smaller LLM. It is a probabilistic decision model: an application sends it a block of program state plus a set of typed questions, and Jev returns structured answers — a Choice among up to 255 options, a Score on an ordered scale, or a Noul, a yes-no probability — each carrying calibrated confidence. Every question in a request is evaluated in parallel rather than generated token by token.

TypeSafe trained it with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD), so that stated confidence tracks actual accuracy: code can act automatically above a threshold and route to a human below it. Schema matching is guaranteed by construction, which means an agent choosing between four tools cannot invent a fifth. Because the model never generates free-form strings, it cannot hallucinate in the way language models do.

The pricing is the other headline: $0.042 per million input tokens, with output tokens free — “too cheap to meter,” in the company’s words, against $0.20 to $10 per million input for frontier LLMs, whose output is typically several times more expensive than input. End-to-end latency runs 70–500ms, versus 3 to 329 seconds for frontier models.

Why adoption happened this fast

The Vercel blog attributes the speed to specialization: a model built for exactly one slot in the stack — the decision layer — finds that slot in production almost immediately. The use cases write themselves: choosing an agent’s next tool or subagent, deciding whether a workflow should continue, retry, ask the user, or stop, scoring urgency or risk before an action, and verifying model outputs or enforcing guardrails before anything autonomous runs.

The surrounding ecosystem moved within days. Vercel added Jev to AI Gateway on September 16 and listed tool selection, next-action choice, risk checks and output validation as target use cases. Cloudflare listed it in its AI docs as typesafe/jev, callable via a plain REST POST that reuses the Workers AI token. LangChain shipped a TypeSafeClassifier integration, and Langfuse published a guide to eval scoring with it.

TypeSafe’s own workflow evaluations — spanning security alerts, agent review, invoice processing and customer service — put Jev at 67.8% agreement with the averaged answers of GPT-6 Astra and Claude Fable 5.1, level with GPT-5.6 Terra and Claude Sonnet 5, at $0.0004 per case and 0.4 seconds. Sonnet 5 reaches the same score in 78.1 seconds at 294 times the cost per case. The vendor claims peak multiples of 193.6x faster and 444.6x cheaper sit at the higher end of real-world gains — a qualification the company publishes itself, alongside the fact that reference labels come from other models’ answers rather than ground truth.

Independent anecdotes are stacking up in the same direction. Every’s Mike Taylor ran 37 documents against 21 writing checks and got 777 judgments back in under 0.7 seconds for about a quarter of a cent. One developer pointed Jev at 9,081 product-matching pairs his own engine had parked for human review and cleared the entire queue for 32 cents in 13 minutes — a stage he had abandoned in June because frontier output pricing made it uneconomical.

The Jevons bet behind the name

The name is not an accident. TypeSafe named the model after William Stanley Jevons, the economist who observed that cheaper coal increased coal consumption. The company expects intelligence to follow the same curve: every order-of-magnitude drop in the cost of decisions unlocks orders of magnitude more use cases. The 32-cent review queue is that effect in miniature — work that did not exist at $15 per million output tokens exists at zero.

DCVC led a $40 million seed round to back founder Diogo Almeida, who helped build the instruction-following research behind ChatGPT at OpenAI before spending two years in stealth with co-founders Erik Gafni and Sasha Sheng. The buyer profile is telling: agent vendors like Decagon, which raised $250 million at a $4.5 billion valuation in January, run support agents whose every routing, scoring and escalation step is currently a language model call. Moving those steps to a model with free output expands margins across the entire agent industry at once — while the lab owning the decision layer sits underneath all of them.

Caveats worth keeping

A model five days old with vendor-run evals deserves skepticism, and TypeSafe mostly agrees: it publishes nuance boxes under its own claims, notes that its workflow evals were built by its own capabilities team, and admits the side-by-side demo inputs paint the model in an advantageous light. LLM numbers in its comparisons come from OpenRouter, which may route harder queries to better models. The long-term sustainability of free output pricing is unproven by definition — only time will test it.

There is also a competitive frontier already: within days of Jev’s launch, open-weight “System 1” decision models such as Laya appeared, claiming 33ms latency and better scores at no API cost at all. If decisions and text really do separate into different model classes — the bet Vercel’s adoption curve is now testing — the decision layer of the agent stack is about to become its own fiercely contested market.

Vercel’s own framing is measured: the next test is whether that early adoption lasts. But a record is a record, and it was set by a model that gave up words entirely.