Jev is not a chatbot. It is TypeSafe AI's first System One model: a decision engine that takes messy app state and returns typed answers your software can act on.
TypeSafe launched Jev in early access on September 14, 2026. If you already use tools like Lovable, Base44, Make, or n8n to build apps with AI, Jev is a different layer. It does not write the app. It sits inside an app you already ship and answers one constrained question at a time.
We will walk TypeSafe's support-ticket example, then show how to try it this week. Speed and price figures below are TypeSafe's published numbers, not independent benchmarks.
What is Jev AI?
System One is TypeSafe's name for models that evaluate a state and return typed answers plus probabilities. Jev is the first public one. The default alias is jev-latest, which currently points to jev-1.13.0.
The name comes from Daniel Kahneman's System 1: fast judgment. Founder Diogo Almeida, who helped invent the instruction-following methods behind modern chat models, argues chat got superhuman while software still does not run itself. Jev is built for decisions inside code, not conversation.
Text only today: a string, a JSON object, or an array of text values. No images, audio, or video. Convert those to text first.
Output is constrained. Jev does not write a reply, generate code, or explain itself. You define the answer space. It returns values your code can branch on, plus probabilities and, for Choice and Score, a confidence score.
Jev vs ChatGPT, Claude, and other LLMs
Same intelligence job, different product. A typical LLM is trained to write. You parse the string, hope the JSON is valid, and retry when it is not. Jev is trained with Reinforcement Learning for Calibrated Decisions (RLCD) to pick from answers you already defined.
Typical LLM: trained for chat and verifiable text. Output is a string you must parse. Sampling is one token at a time. Latency is often seconds. You pay for input and output tokens.
Jev / System One: trained for calibrated decisions. Output is a typed value. Independent questions run in one parallel call. TypeSafe's docs describe roughly 100 ms per query. Official pricing as of this writing is $0.042 per million input tokens, with output tokens free. See the models page for the current table, rate limits, and aliases.
TypeSafe's speed and cost claims are company numbers. We did not independently benchmark Jev. Typed output is also not the same as a correct output. Confidence is there so your app can act, ask a person, or fall back to a chat model.
The three questions you can ask: Choice, Score, and Noul
You do not prompt a novel. You ask a typed question. TypeSafe currently ships three primitives, documented under Primitives (Questions): which option, how much, and yes or no.
Cheat sheet. Same messy input, three jobs. Extra examples are question shapes, not live Jev calls.
| Type | Question it answers | Example you would ask | What comes back | What your app does |
|---|---|---|---|---|
| Choice | Which of these options? | Which team owns this ticket: billing, technical, or sales? | One option, a probability for each option, and confidence | Route the ticket to that queue |
| Score | Which level on a scale you wrote? | How frustrated is this customer: calm, frustrated but civil, or very angry? | A level, a probability for each level, and confidence | Escalate if the score is high |
| Noul | Is this true? | Does this message request a refund? | A number from 0 to 1 (probability of yes) | If noul is high, start the refund flow |
| Choice | Which of these options? | What kind of form is this: demo, support, partnership, or spam? | One option plus probabilities and confidence | File the row in the right Airtable or CRM view |
| Score | Which level on a scale you wrote? | How ready is this lead: browsing, comparing vendors, or ready to buy? | A level plus probabilities and confidence | Send ready-to-buy leads to sales now |
| Noul | Is this true? | Did the other model follow the policy before this reply goes to a customer? | A number from 0 to 1 | If noul is low, hold the reply for a human |
Choice: pick one option from a menu you already wrote
Use Choice when you can write the answers before you see the input: billing, technical, or sales; demo, support, partnership, or spam; search, calendar, send_email, or do nothing. Jev picks one.
You get the selected option, a probability for every option, and a confidence score. Your Make scenario, n8n workflow, or app then branches. No 'please respond with only the category.' No JSON repair loop.
Score: rate it on a scale you defined
Use Score for how much, not which bucket. Do not ask for a free-form 1 to 10. Write ordered levels in plain language, and Jev places the state on that scale.
Frustration is the official example: calm, frustrated but civil, or very angry. The same pattern works for lead readiness (browsing, comparing, ready to buy) or for a guardrail on another model's draft (safe, needs review, block).
You get a score, a probability for each level, and confidence. Escalate at a level you chose. The model does not send the angry email. It tells you the customer sounds angry.
Noul: yes or no, as a probability
Noul is TypeSafe's made-up name for yes/no. The answer is not the word yes. It is a number from 0 to 1: P(yes). Your code picks the threshold.
Good Noul questions are snap judgments: does this request a refund, look like a real person, or follow the policy? TypeSafe's docs: ask for a judgment a knowledgeable person could make in a second.
If noul is 0.96, you can auto-act. If it is 0.51, do not pretend that is a yes. Treat the number as P(yes) and let software decide.
You can ask several independent questions in one call. Jev scores them in parallel against the same state. TypeSafe says adding questions barely changes response time. If a judgment has three unrelated factors, ask three questions.
Walkthrough: one support ticket, three decisions
TypeSafe's official quick start uses a Stripe ticket. A customer writes: Hi, I've been trying to connect my Stripe account for 3 days and it keeps failing. I'm losing sales. Please help ASAP.
A chat model would start drafting an apology. Jev does not. Send that text as state and ask all three at once:
Choice: which team should handle this (billing, technical, or sales)?
Score: how frustrated the customer appears (calm, frustrated but civil, or very angry)?
Noul: does the message convey urgency?
TypeSafe's published sample, not a live call we ran, returns roughly: department = billing at 0.84 with confidence about 0.60; frustration about 1.0 on a 0-2 scale; urgency noul = 0.999. Your code can route, escalate, or hand a chat model a brief to write the email.
Chat models still write. Jev decides. The answer says what; confidence says whether to act or escalate.
Why builders should care
Put Jev where handwritten rules get brittle and a full agent is overkill.
Smart if-statements. Classify, route, score, or extract inside an otherwise normal workflow. A form submission, a help-desk ticket, a lead record, or a row in Airtable is enough state.
Confidence gates. Act when confidence is high. Escalate to a person or a chat model when it is not.
Guardrails. Score, judge, or check other LLM outputs before they hit a customer. If you already give agents tools through something like MCP servers, Jev is a narrower job: one typed check, not a new agent runtime.
Cost. Official list price is $0.042 per million input tokens, output free. Rate limits on the models page are currently listed as 250,000 tokens per second and 1,200 requests per minute, and TypeSafe says those can change while demand is high. Higher limits are listed as custom/enterprise. Context is 64k tokens per request.
How to use Jev this week
Do not rebuild your stack. Add one decision step.
1. Playground. Log in at the TypeSafe console. Paste a real ticket or email as state. Add one Noul, then mix Choice and Score. If you do not have access yet, TypeSafe still has a waitlist. Early access is not the same as a fully open public API.
2. HTTP API. POST https://api.typesafe.ai/v1/systemone with model jev-latest and an API key from the dashboard. Send state plus a questions object. The response includes answers, usage, and the versioned model ID that actually ran.
3. SDK or skill. Python package typesafe-sdk, JavaScript client @typesafe-ai/sdk, or the TypeSafe agent skill for Claude Code and other coding agents. The SDK default is jev-latest.
4. No-code path. Wrap the same HTTP call as one module in Make, Zapier, or n8n, or as a backend function in an AI app builder. Pattern: inbound lead or ticket -> Jev scores intent, urgency, and routing -> your automation branches -> Claude or GPT writes the human-facing message. Keep chat models for copy. Keep Jev for the judgment.
When not to use Jev
Skip Jev if the job is writing: blog posts, emails, UI copy, or tutorials. Use a chat model.
Skip Jev if the job is generating an app or a workflow. That is still vibe coding and no-code app builders. Jev sits inside software you already control. It does not pick the next agent step for you.
Skip Jev if the input is a screenshot, audio, or video and you have not turned it into text.
Use Jev when you can name the answer in advance, when you want a probability instead of a paragraph, and when your code (or your Make scenario) should stay in charge of what happens next.
Bottom line
Unstructured state in. Typed decisions out. Jev will not replace Claude for a landing page, and it is not a smaller ChatGPT. Open the playground with one real ticket, ask Choice, Score, and Noul, then decide if a decision step belongs in your app.
Sources: TypeSafe launch post, System One docs, quick start, models and pricing, and how to build with System One. All product facts in this article were checked against those pages on September 18, 2026.
Frequently Asked Questions
What is Jev AI?
Jev is TypeSafe AI's first System One model. It takes unstructured text state and returns typed Choice, Score, or Noul answers with probabilities so software can branch without parsing a chatbot reply. It launched in early access on September 14, 2026.
What is the difference between Choice, Score, and Noul?
Choice picks one option from a set you define, such as billing, technical, or sales. Score rates a state on ordered levels you write, such as calm, frustrated, or very angry. Noul returns a number from 0 to 1: the probability that a yes/no question is true. You can ask several in one call against the same state, then compose the answers in your app.
How much does Jev cost?
TypeSafe's models page lists Jev 1.13 at $0.042 per million input tokens, with output tokens free. That is company pricing as of September 18, 2026, not an independent cost study. Rate limits and aliases can change.
Is Jev a chatbot or an agent?
No. Jev does not write replies, generate code, or choose the next step in an agent loop. You define the answer space. Your app keeps control of workflow, tools, and side effects.
Can I use Jev in a no-code app?
Yes, as an HTTP decision step. Call POST https://api.typesafe.ai/v1/systemone from Make, Zapier, n8n, or a backend function in an AI app builder. Use Jev to classify or score; keep a chat model for any customer-facing writing. Access is still early access, so you may need a TypeSafe account or waitlist approval first.
What is System One?
System One is TypeSafe's class of models for fast, structured decisions. Jev is the first public model in that class. Inputs are text; outputs are constrained typed values plus probabilities, trained with Reinforcement Learning for Calibrated Decisions (RLCD).
Advertiser disclosure: some links on this website are affiliate links, meaning No Code MBA may make a commission if you click through and purchase.