What Is Jev? TypeSafe AI's Decision Model Explained

Jev is an AI model from TypeSafe AI that makes decisions instead of writing text. You give it some information and a list of questions with fixed possible answers. It returns an answer to each question with a probability attached, usually in 70 to 500 milliseconds. It launched in limited early access on 15 September 2026 and costs $0.042 per million input tokens, with output not charged at all.

TypeSafe calls Jev a "System One" model. The name comes from psychology, where System One is the fast, automatic kind of thinking and System Two is the slow, deliberate kind. ChatGPT, Claude and Gemini are built to do the slow kind. Jev is built to do the fast kind, very cheaply, millions of times a day.

This guide draws on TypeSafe's launch coverage in TechCrunch, DataCamp's technical explainer, LangChain's integration guide and Wikipedia, all linked at the bottom. Speed and cost comparisons are TypeSafe's own claims. Jev is waitlist-only, and we have not used it on client work yet.

Jev at a glance

Almeida previously worked at OpenAI on the research behind ChatGPT, according to TechCrunch. The model is named after William Stanley Jevons, the 19th-century economist who noticed that making something cheaper to use tends to increase how much of it gets used.

How Jev works

With a normal AI model you send a prompt and it writes an answer one word at a time. You then hope it is in the format your software expects.

With Jev you send two things:

1. The state. Whatever you want it to judge. An email, a form submission, a product review, a log of what an AI agent just did.
2. The questions. Each one has a type and a fixed set of possible answers.

There are three question types:

• Choice picks one option from a list you define, with a probability for each. For example: is this enquiry about pricing, booking, a complaint or something else?
• Score rates something on a scale you define. For example: how urgent is this, from 0 to 100?
• Noul gives the probability that a yes/no statement is true. LangChain's guide shows the statement "The message conveys urgency or time-sensitivity" returning 0.999.

Jev answers all the questions in one pass, in parallel, rather than writing them out. That is where the speed comes from.

TypeSafe trained it on synthetic data using a method it calls Reinforcement Learning for Calibrated Decisions. The aim is that the confidence scores mean what they say: if Jev is 90% sure across a hundred decisions, it should be right about 90 times.

Why it cannot make things up

The most common complaint about AI in business is that it confidently invents things. Jev cannot do that in the usual sense, because every answer has to be one of the options you gave it. It can still pick the wrong option. What it cannot do is invent an option you did not offer, or hand back something in the wrong format and break the next step of your automation.

DataCamp's explainer reports a test comparing Jev with two general-purpose models on the same classification cases:

That table is worth reading carefully. Jev was not the most accurate. Claude Opus 5 was, by about five points. But Jev matched GPT-5.6 Terra on accuracy at roughly a seventy-fifth of the cost and a twenty-fifth of the time, and never broke format. For a decision you need to make ten thousand times a day, that trade is often the right one.

What it costs

At $0.042 per million input tokens with free output, Jev is priced in a different range from the models it is compared with. For comparison, Claude Opus 5.5 is $4 per million input tokens and $20 per million output, and GPT-6 Astra is $10 and $50.

A 2,000-token email costs under a hundredth of a US cent to classify. A thousand of them costs about 8 cents. The same thousand written up by Opus 5.5 would be around $28, because Opus 5.5 is also writing a reply, which Jev never does.

TypeSafe claims Jev is up to 200 times faster and up to 400 times cheaper than comparable LLMs on classification tasks. Those are the company's figures and they depend heavily on what is being compared, so read them as "very large", not as exact.

Where Jev fits in a business

Jev is not something a business owner would use directly. There is no chat window. It sits inside an automation and makes the small calls that currently either need a person or an expensive AI model.

Sorting enquiries. Every form submission or email gets tagged by type and urgency before anyone sees it. Urgent complaints go straight to the owner's phone. Pricing questions get an instant, accurate reply. Spam never reaches the inbox.

Scoring leads. A new enquiry comes in and Jev scores how likely it is to be a good fit, based on the rules you set. The high scorers get a same-day call.

Routing to the right AI model. LangChain's guide describes using Jev to decide whether a task needs a powerful model or a cheap one, so you only pay for the expensive model when it is actually needed.

Checking another AI's work. Before an AI agent sends an email or runs an action, Jev can check whether it looks risky and block it. LangChain describes exactly this, using Jev to check an agent's tool calls before they run.

Moderating reviews and messages. Flagging reviews that break a policy, or customer messages that need a human, at a volume no one could read manually.

This is the layer we already build into AI automations for clients, usually with a general-purpose model doing the sorting. Jev would do the same job faster and for less once it is openly available.

The limitations

It cannot write anything. No emails, no summaries, no explanations. If a step needs words, you still need a model like Claude or ChatGPT.

It does not explain itself. You get a probability, not a reason. If you need to tell a customer or a regulator why a decision was made, you will need something else alongside it.

You have to define the answers up front. Jev is only as good as the options you give it. Messy, open-ended questions are a job for an LLM.

Little is published about how it is built. TypeSafe has not released its architecture, weights or a technical paper, according to Wikipedia. Independent testing is still thin.

It is waitlist-only. As of launch, developers are being let in from a waitlist over time. TechCrunch reports demand was high enough that the API was briefly unavailable.

Should you care about Jev?

If you run a small business, not directly. If you use, or are thinking about, automations that sort and route things (enquiries, bookings, support messages, reviews), then yes, because it points to where the cost of that kind of automation is heading. The expensive general-purpose models will do the thinking and the writing. Cheap, fast models like Jev will do the thousands of small decisions around them.

For a side-by-side of how Jev compares with Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra, see our AI model comparison.

Sources

• A new kind of AI model from a ChatGPT inventor is thrilling developers, TechCrunch
• Jev: TypeSafe's System One Model Explained, DataCamp
• Building a harness with Jev, LangChain
• ChatGPT pioneer launches Jev model for programmatic logic, AI News
• Jev (AI model), Wikipedia

Last checked against these sources on 22 September 2026.