[ What is Jev? ]
What is Jev?
The decision model, explained.
Updated 1 October 2026 · by JevHub
Built to decide, not to write
Language models like GPT and Claude generate text one token at a time. That makes them flexible, and also slow and expensive when all you need is a decision.
Jev drops text generation entirely. You send an input (an email, a ticket, a lead record) and one or more typed questions. It evaluates them in parallel and returns structured answers: the option it chose, a probability for each option, and a confidence score. TypeSafe describes it as a function call with frontier-level judgment: unstructured input in, typed decisions out (announcement (opens in a new tab)).
That changes what is worth automating. One builder triaged 1,000 emails in about 1 minute for 3¢. Another classified 1,018 research papers for 8¢. At those prices you can run a decision on every item instead of sampling.
Fig.1 · Same question, two ways
The question
“Is this email urgent?”
A language model
GPT, Claude, Gemini
A paragraph, generated one token at a time.
- Output
- text, one token at a time
- Latency
- seconds
- Next
- parse and validate the text
Jev
Chooses from the options you define
- Output
- a typed answer with probabilities
- Latency
- 70–500 ms
- Next
- your code acts on it
Inputs and outputs
A request has two parts: the input to evaluate and the questions about it. Each question is one of three types (Choice, Score or Noul) and declares the answers it accepts, so a response can never fall outside them.
Questions about the same input are answered in a single request, so asking three costs little more than asking one.
Confidence is the point
Every answer comes with a confidence score, and that is what makes automation safe. Set a threshold: above it, act automatically; below it, escalate to a person or a larger model. Clear cases are handled in milliseconds, and people only see the ambiguous ones.
Fig.2 · Confidence thresholds
Below thresholdEscalate to a person or a larger model
AboveAct automatically
TypeSafe says Jev is trained for calibration, meaning higher confidence should correspond to higher accuracy (announcement (opens in a new tab)). Validate that on your own data before relying on it.
[ How it works ]
Three question types.
Every answer comes with a probability.
Jev doesn't generate text. It evaluates your input against the questions you define and returns typed answers with calibrated probabilities.
Fig.1 · Choice
Choice: one option from a list
Route a ticket, label an email, pick the next action. You define the options.
Fig.2 · Score
Score: a position on a scale
Urgency, lead fit, quality. Anything with a natural order from low to high.
Fig.3 · Noul
Noul: a yes/no probability
Is it spam? Does it need a human? Built for gates and flags.
It complements language models
Few teams replace an LLM with Jev. They run both: Jev handles the high-volume judgments (route, score, flag, verify) and a language model handles the work that needs words. Common pairings are triage before generation, guardrails on model output, and routing requests between models.
| Jev | Language models (GPT, Claude, Gemini) | |
|---|---|---|
| Output | A typed decision with probabilities and a confidence score | Generated text |
| Cost | ≈ 2.4¢ per 1,000 emails | 5–8× more in one independent test, even for small models |
| Latency | 70–500 ms | Seconds |
| Writing, reasoning, chat | No | Yes |
The cost comparison comes from an independent test; see costs for the details.
Limits
Worth knowing before you build:
- No text generation. No replies, summaries or code. That stays with a language model.
- Closed answer sets. It chooses only from the options you define, so you need to know them in advance.
- No explanations. You get the decision and its confidence, not the reasoning.
- Text only. No images, audio or video.
- It can still be wrong. The answer is always a valid option, but not always the right one. That is what the confidence score is for.
Who built it, and where the name comes from
Jev was released on 15 September 2026 by TypeSafe AI. Its founder, Diogo Almeida, worked at OpenAI on the research behind ChatGPT, then spent two years in stealth building Jev (TypeSafe's announcement (opens in a new tab)).
The name comes from the Jevons paradox, according to his launch post (opens in a new tab). In 1865 the economist William Stanley Jevons observed that as steam engines burned coal more efficiently, Britain burned more coal, not less, because engines became worth using everywhere. The bet behind Jev is the same: make a decision cheap enough, and software will make millions of them.
TypeSafe also calls Jev a “System One” model (docs (opens in a new tab)), after Daniel Kahneman's Thinking, Fast and Slow: System 1 is fast and intuitive, System 2 slow and deliberate. Text-generating models are the deliberate kind. Jev is built for the fast kind.
[ Getting started ]
Three ways in.
Try it, use it, or build it.
[ Try it ]
Test it in the playground
Paste an input, define a question and its options, and see the answer with its probabilities. No code required.
Open the playground (opens in a new tab)[ Use it ]
Use an existing tool
Builders have already shipped tools on Jev, including an open-source inbox triage app. See what exists before building.
Ready-made tools[ Build it ]
Integrate it
Call Jev from your inbox, CRM or help desk. Each use-case page includes the exact questions and the request.
Browse use cases[ Questions ]
Common questions
- What is Jev?
- Jev is a decision model from TypeSafe AI, released on 15 September 2026. It doesn't generate text. You send an input and typed questions, and it returns one of the answers you defined, with a probability for each option and a confidence score.
- Is Jev like ChatGPT?
- No. GPT, Claude and Gemini generate text. Jev only returns structured decisions. That makes it far cheaper and faster for classification, scoring and routing, but it can't write replies, summaries or explanations.
- How much does Jev cost?
- $0.042 per million input tokens, with output free. Triaging 1,000 emails costs about 2.4¢ at list price. Use the cost calculator for your own workload.
- Do I need to code to use Jev?
- Not to evaluate it: TypeSafe's playground runs in the browser. To run it on live data you call its API, or use a tool someone has already built on it.
JevHub is independent and not affiliated with TypeSafe AI.