Jev AI Decision Model: What It Does and What It Costs

Platinum Web Studio scored all 60 pages of its own website with Jev, one call per page and six questions per call, and the Vercel AI Gateway reported a market cost of about $0.011 for the entire run. Jev, released by TypeSafe on September 15, 2026, does not write, chat, or summarize. It answers typed questions about a piece of text with a typed value and a confidence figure, which makes it a decision engine rather than a writer. This guide is for Los Angeles service business owners, from HVAC and plumbing companies to appliance repair shops, who are weighing whether AI automation should sort their customer inquiries.

What the Jev AI Decision Model Actually Does

TypeSafe calls Jev a System One model, borrowing the psychology term for fast, intuitive judgment. A chat model such as Claude or GPT generates its answer one token at a time. Jev takes two inputs instead: a state (the text you want judged, such as an email, a contact form, or a web page) and a list of typed questions about that state. It evaluates every question against the state in parallel and returns structured values your code can use directly, each with a probability and a confidence figure attached.

TypeSafe's documentation names three question types:

• Choice. Pick one option from a list you define, for example “AC repair, AC install, plumbing, or not a service request.” The answer includes a probability for every option.
• Score. Rate the state against ordered levels you describe in plain language, for example five levels of how complete a quote request is.
• Yes or no. TypeSafe calls this a noul: a statement such as “this message asks for a price on a specific job,” answered with the probability that it is true.

You write the categories and criteria. Jev never invents its own labels, so its output can feed a routing rule or a CRM field directly.

How Fast and How Cheap Is Jev? TypeSafe's Own Numbers

Every speed and cost multiplier in circulation comes from TypeSafe itself, not from independent testing. The launch post by founder Diogo Almeida, who says he helped build the methods at OpenAI that made language models follow instructions and hold a conversation, states that Jev answers end to end in 70 to 500 milliseconds and that this can be 40 to 200 times faster than frontier models at the same level of intelligence. TypeSafe's homepage goes further, claiming 193.6 times faster and 444.6 times cheaper, and the launch post says those figures come from TypeSafe's own workflow evaluations on System One tasks.

The pricing is easier to check because it is published. TypeSafe's model page lists Jev 1.13 at $0.042 per million input tokens ($42 per billion) and makes output tokens free, since the answer is a short typed value rather than paragraphs of text. You pay for what you send, not for what comes back. Treat the multipliers as the maker's claims until your own task, text, and questions confirm them.

What We Learned Scoring 60 Pages With Jev

We did not test Jev on a demo. We used it to grade our own site on six dimensions: search intent, spec compliance, factual density, originality, conversion clarity, and link health. Each question was a score on a 1 to 5 scale, with written criteria for every level.

Scores were stable on identical copy

Re-scoring pages whose copy had not changed moved most scores by 0.0 to 0.1 on the 1 to 5 scale, and the largest single move on any page was 0.3. That is stable enough to compare a page before and after a rewrite, as long as the change you are measuring is bigger than a few tenths of a point.

Jev judged only what it was given

An earlier run cut each page's text at 5,000 characters to keep calls small. On 3 pages the FAQ sat past that cutoff, so Jev never saw it and recommended adding an FAQ that already existed. The model was right about the text it received; the input was wrong. Sending the full page text fixed it.

What Jev Cannot Do

Jev cannot write a reply, summarize a call, draft a quote, or hold a conversation. If a task ends in sentences a customer reads, Jev is at most the first step.

Its input has limits too. TypeSafe's documentation sets Jev 1.13 at 64k tokens per request, with 32k of that for the state plus the longest single question, and accepts text only: no images, audio, or video. TypeSafe also says English is where accuracy is currently best, so a voicemail or a Spanish-language text needs converting or extra checks first. And a typed answer is only as good as the question: TypeSafe's guidance is to ask one specific, well-scoped thing per question and combine the answers in code.

Using an AI Decision Model for Inquiry Triage: A Real Test

Most of what a contractor's inbox needs is a decision, not a paragraph. We tested exactly that case. Jev read two real inbound messages with one yes or no question each. It classified a white-label web design agency's cold pitch as outreach and a real plumber's quote request as a genuine customer inquiry.

That single question is the core of inquiry triage: a real request buried behind vendor pitches can cost a job. A rule like this is part of what we build into a pipeline through our CRM and customer follow-up service, so each new form is tagged as a real request or outreach before anyone opens it.

Other Everyday Decisions Jev Can Handle

The same pattern covers most of the daily sorting a service business does by hand:

• Service type. A Choice question sorts each request into the jobs you actually do.
• Urgency. A Score question rates how urgent a message reads (“no heat, two kids at home” versus “quote for next spring”), so the urgent one is called first.
• Review tone. A Score question grades each new Google review, so an angry one-star review gets a same-day reply from the owner instead of sitting unread. Fast, personal replies are a core part of how our online reputation management service handles negative reviews.
• Spam. A yes or no question filters bot submissions before they reach the CRM.

TypeSafe's documentation calls this intent routing: classify each request, then send it to plain code, a specialist language model, or a person.

How Platinum Web Studio Uses Decision Models in Inquiry Handling

When we set up inquiry handling for a service business, the decision model is one step, never the whole system. It sits between the inquiry source and the person who calls back:

1. Each new form, text, or call transcript gets the same short list of typed questions: real request or outreach, which service, how urgent.
2. Plain rules act on the answers: urgent requests reach the owner first, vendor pitches are tagged and set aside.
3. Anything below the confidence line waits for a person, and every reply a customer reads is still written by a person or a language model.

In the free 15-minute audit, we look at your website and Google Business Profile with you and show where calls and form requests are leaking. Bring your most repetitive sorting task and we will tell you whether a decision model can take it or whether it should stay with your team. You leave knowing what to fix first, with no pitch; if you go ahead, the next step is a confirmed scope and price before anything is built.