What Is Jev AI? TypeSafe's Decision Model Explained

What Jev is

Most AI models released in the last few years do the same thing: you send them words, and they send words back. Jev does not.

Jev launched on 15 September 2026 from TypeSafe AI, a startup that came out of stealth that day with a $40M seed round led by DCVC. It was founded by Diogo Almeida, a former OpenAI researcher and co-author of the InstructGPT paper, one of the foundations of ChatGPT, with co-founders Erik Gafni and Sasha Sheng. The name is a nod to the Jevons paradox: when something gets much cheaper, people end up using far more of it.

The pitch is simple. Jev is a model for machines, not people. You do not chat with it. Your software asks it a question with a fixed set of possible answers, and Jev returns the answer with a probability for each option.

TypeSafe calls this a System One model, after Daniel Kahneman's split between fast, intuitive judgement (System 1) and slow, deliberate reasoning (System 2). Today's language models behave like System 2: they think step by step, write out their work, and take seconds to respond. That is overkill when the only question is whether an email is a complaint or a sales enquiry.

How Jev works

A Jev call has two parts. The state is the context you want judged: text, structured data, or an agent's message history. The questions are what you want to know about it, each with a predefined answer format.

There are three types of question:

Here is how a single support message might be judged in one call. The answers are illustrative, not real output:

There is no text to parse and no chance of an answer outside the options you set.

Ask everything at once. Jev evaluates every question in a request in parallel, against the same state. Extra questions only add their own tokens, so the recommended pattern is to ask everything you might need in one call. In one published test, batching 13 questions into a single request was about 12 times cheaper and 10 times faster than asking them one by one.

The probabilities are meant to be honest. TypeSafe trains Jev with a method it calls Reinforcement Learning for Calibrated Decisions. The aim is that when Jev says it is 90% sure, it is right about 90% of the time. That matters because the probability is what you build rules on: auto-approve above 0.95, for example, and send everything else to a person.

Can Jev really not hallucinate?

TypeSafe advertises "zero hallucinations". It is worth being precise about what that means.

What is true. Jev can only answer with options you defined in advance. It cannot invent a source, make up a URL, or return malformed data. In TypeSafe's own tests it had a 0% error rate on structured output and tool calls, where the language models it was compared against did not.

What is not. Jev can still be wrong. If the right answer is Billing and it picks Sales, that is a mistake, even if it is not a hallucination in the usual sense. It does not invent things. It just picks the wrong option.

How good is it?

TypeSafe tested Jev against three frontier models on four production-style workflows. These are the vendor's own numbers and have not been independently verified.

That is the trade-off. You give up some accuracy for a very large drop in cost and wait time. For a million routine decisions a day, the trade usually makes sense. For one high-stakes decision, it usually does not.

Where Jev fits in an AI agent

An AI agent is usually a language model in a loop: think, pick a tool, run it, look at the result, repeat. Around that loop sits the harness, the code that decides what the agent is allowed to do, when it has finished, and when something has gone wrong.

That harness is full of small decisions. Today most of them either run on the same expensive model or are not made at all. These are the ones Jev is designed for:

1. Routing requests to the right model

   Decide whether a request is simple enough for a cheap, fast model or needs the most capable one.

   This is one of the two patterns LangChain's guide to Jev builds out, and the one with the clearest cost saving.

   Question type: choice

2. Checking actions before they run

   Before the agent deletes a file, runs a shell command or sends an email, ask whether the action is risky. Block it or ask a person if it is.

   Jev is not hardened against deliberately hostile input yet, so this should be one layer of protection, not the only one.

   Question type: yes/no

3. Choosing the right tool

   Pick the best tool for the next step from a catalogue. Choice questions support up to 255 options.

   Question type: choice

4. Knowing when the job is actually done

   Agents often say a task is finished when it is not. A cheap completion check on every run catches this before a person has to.

   The same approach spots loops, where the agent keeps repeating the same step.

   Question type: yes/no

5. Deciding when to escalate

   Score how confident the agent should be and hand anything below your threshold to a person, with the probability attached.

   Question type: score

6. Grading runs at scale

   Score thousands of past agent runs to find where they fail. At a fraction of a cent per check, scoring everything becomes realistic rather than sampling.

   Question type: score

The point is that Jev does not replace the model that runs your agent. It sits next to it. The language model does the thinking and writing, and Jev makes the fast calls along the way.

The same pattern works outside agents. A shared inbox is a good example: one Jev call per message could tag the topic, score urgency and flag whether it needs a senior person, then a person or a language model writes the reply. That is the routing automation most small teams want, with the judgement step handled cheaply.

What Jev cannot do

According to TypeSafe's own guidance and early reviews, Jev:

• Does not generate text. No writing, code, summaries or explanations.
• Does not explain its decisions. You get a probability, not a reason, which is a real problem where decisions have to be justified or audited.
• Is weak with numbers and dates. It cannot reliably count, do arithmetic, compare numbers or reason about dates. Do that in code first and pass Jev the result.
• Struggles with tricky wording, such as double negatives and indirect references.
• Cannot extract values. It picks from your options. It cannot pull out an arbitrary name or amount.
• Reads text only. No images, audio or video yet.
• Needs design up front. You define the questions, options, thresholds and escalation paths before you start.
• Comes from a very new vendor. It is in early access and, according to InfoWorld, hosted in a single region, from a company that came out of stealth in September 2026.

There is also a fair question about how long the advantage lasts. Some engineers argue that existing language models, limited to short structured answers, could get much closer to Jev's speed than the headline figures suggest. And a model that skips step-by-step reasoning will likely stay behind the best models on hard judgement calls.

Pricing and access

Prices and limits are as reported at launch on 15 September 2026 and may change.

Jev is not a smarter chatbot and is not trying to be one. It makes the many small, repeated decisions inside software fast, cheap and structured. For teams running AI agents, that means checks that used to be too slow or expensive to run on every step now cost almost nothing, as long as the language model keeps the thinking, people keep the high-stakes calls, and nobody reads "cannot hallucinate" as "cannot be wrong".

Sources: TypeSafe launch announcement, LangChain, Flavio Copes, DataCamp, InfoWorld, The Rundown AI, heise online, Sean Goedecke and TypeSafe AI.

Frequently asked questions