What is Jev, the decision model from TypeSafe AI?

What is Jev?

Jev is TypeSafe AI’s decision model. You send it some content, called the state, plus typed questions. It returns one structured answer per question, with probabilities. It generates no text at all.

TypeSafe calls it the first System One model. The name comes from Kahneman’s System 1, the mode of fast judgments. We cover the concept in our article on System One models.

In practice, Jev replaces the LLM call you ask to “reply in JSON” so you can sort, classify or score something. Your code gets back a value it can drop straight into an if.

Who makes Jev?

• The company: TypeSafe AI, based in San Francisco and founded in 2024.
• The launch: Jev launched with a waitlist on 15 September 2026, with a launch post from TypeSafe. It opened to everyone on the 20th, then new sign-ups were paused on the 22nd because of demand.
• The funding: a $40M seed round led by DCVC, according to Wikipedia and press coverage.
• The founders: Diogo Almeida, formerly at OpenAI, with Erik Gafni and Sasha Sheng. TypeSafe describes Diogo Almeida as a co-inventor of RLHF, the training method behind ChatGPT.

What does Jev return?

Jev supports three question types, which TypeSafe calls primitives. You can mix them in a single call.

• Probabilities for a choice add up to 1.
• Confidence summarises how they are spread. The docs give an approximation, (n × p − 1) / (n − 1), where p is the top probability. Our real calls match it: with 3 options and 0.75 on the winner, Jev returned a confidence of 0.63.
• Score is the probability-weighted mean of the levels. It can land between two levels.
• Noul has no confidence field. You set the threshold yourself.
• The model field tells you which exact version answered.

Here is what a request looks like. Each key under questions is a name you choose. The answer comes back under the same key.

{
  "model": "jev-1.13.0",
  "state": "Hi, I was charged twice this month. Can you refund me?",
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this ticket?",
      "criteria": {
        "billing": "Payment or subscription issues",
        "technical": "Bugs or integration problems",
        "sales": "Pricing or account questions"
      }
    },
    "urgent": {
      "type": "noul",
      "instructions": "The message conveys urgency or time-sensitivity"
    }
  }
}

In our test, a French billing ticket sent through our demo relay came back with:

• Team: billing, with a confidence of 1.
• Urgency, as a noul: 0.07.
• Frustration, as a score: 0.02.
• 494 ms end to end, 499 input tokens for 3 questions, answered by jev-1.13.0.

One ticket says nothing about accuracy. It shows the shape of the answers and a real-world latency, relay included. You can run your own try on our demo.

What is Jev for?

Jev is built for closed, repeated, high-volume decisions. The kind where the answer fits in a list you know in advance.

• Routing a ticket, an email or a request to the right team.
• Rating urgency, risk or quality on a rubric you define.
• Checking that a statement holds for a document before acting on it.
• Filtering RAG passages before they reach an LLM.
• Acting as a guardrail on what goes into and out of an LLM.
• Picking the right value among candidates found by a regex.

Confidence is what makes this work. Above a threshold, your code acts. Below it, the case goes to an LLM or a person. That is the pattern in our Jev in n8n guide.

What Jev is not for

• Writing: no customer replies, no summaries, no code.
• Powering a coding agent: the docs rule this out plainly.
• Arithmetic: counting and comparing amounts or dates are documented weak spots. Keep that work in code.
• Multi-step reasoning: double negatives and multi-hop questions lose reliability.
• Anything but text: no images, audio or video.
• Resisting injection: content written to steer the model can move its answer. TypeSafe says so in its list of jev-1.13 weak spots.

How much does Jev cost?

Jev costs $0.042 per million input tokens. Output tokens are free. These are the public prices on the Models page.

Our test ticket used 499 tokens. A million calls of that size would cost about $21.

• TypeSafe warns that rate limits may change without notice while it scales up.
• TypeSafe admits it cannot prove the price is not subsidised.
• Cases you hand back to an LLM are still billed at LLM prices.

Versions and aliases: jev-latest or jev-1.13.0?

Pin the version in production. The only current model is jev-1.13.0. Two aliases point to it.

• jev-latest tracks the latest stable release. It is the SDK default.
• jev-preview tracks the latest release, official or not. There is no preview build yet.

An alias moves when a new version ships. Your answers then change with no change on your side, and so do your confidence thresholds. Log the exact version with every decision.

Does Jev work in languages other than English?

Yes, but English comes first. TypeSafe says English is the primary training language. Other languages are accepted, with lower accuracy.

Our French ticket was classified correctly. That is one case, not a measurement. Before routing non-English traffic, measure Jev on a few hundred of your own cases and watch the confidence.

Where does your data go?

• Hosting: in the United States, according to TypeSafe’s privacy policy.
• Transfers: TypeSafe does not appear on the EU-US Data Privacy Framework list as of 23/09/2026. EU transfers rely on the standard contractual clauses in its DPA.
• Training: Jev is not trained on customer requests. Every account uses the same weights, with no fine-tuning on your data.
• Retention: zero data retention is offered to enterprise customers, per the Legal page.

For personal data, pseudonymise before sending and get the transfer signed off by your DPO.

Limits to weigh before you commit

• Early access: the product shipped on 15 September 2026. Sign-ups are temporarily paused because of demand.
• Vendor numbers: TypeSafe claims up to 193.6 times faster and 444.6 times cheaper. Those figures come from its own workflows, and TypeSafe itself places them at the high end of real-world gains.
• “No hallucination”: that is TypeSafe’s pitch. Jev cannot step outside the schema, but it can be confidently wrong.
• Calibration: it is judged over groups of predictions, never on a single answer. It still needs checking in your language and on your data.

How can you try Jev?

1. TypeSafe’s Playground, if your account is open: paste some text, add questions.
2. The API: POST https://api.typesafe.ai/v1/systemone, with your key in the Authorization header. SDKs exist for Python and JavaScript. See the quick start.
3. Our demo, which calls Jev through our relay.
4. Our n8n guide, for a complete routing workflow with fallback to your LLM.

If an LLM already makes these decisions for you, the real question is different. Does Jev do as well on your cases, for less? That can be measured in shadow mode, without touching production.

Jev and TypeSafe are trademarks of TypeSafe AI, Inc. This article is independent and not affiliated with TypeSafe AI.