What Is Jev? TypeSafe's System One Model, API & Pricing
What is Jev AI? TypeSafe's System One model explained: how it works, pricing ($0.042 per million tokens), API and OpenRouter access, limits and alternatives.
Jev is a hosted AI model from TypeSafe AI and the first System One model. Instead of writing text, Jev reads a piece of input (the state), answers a set of typed questions about it, and returns structured answers with probabilities that your code can use directly: pick an option, score on a scale, or give the probability that a statement is true.
This page explains what Jev is, how it works, what it costs, how to call the Jev API (directly, through OpenRouter or through Vercel AI Gateway), where it falls short, and which open alternatives exist. laya-ai.com is an independent site and is not affiliated with TypeSafe; everything below is based on TypeSafe's public documentation and our own tests.
| Maker | TypeSafe AI |
| Model type | System One decision model (hosted, weights not public) |
| Current version | Jev 1.14 (jev-1.14.0, released October 14, 2026), the model behind jev-latest and jev-preview; Jev 1.13 (jev-1.13.0) is still available (models page) |
| Question types | Choice, Score, Noul; Box (where something is in an image, Jev 1.14 only) |
| Price | $0.042 per million input tokens; output tokens are free |
| Context | 64k tokens per request; 32k for the state plus the longest question. The opt-in TypeSafe-Danger-Zone-Long-Context: true header raises both to 262k, slower and less accurate |
| Input | Text: a string, a JSON object or an array of text values. PNG and JPG images are experimental and need the TypeSafe-Danger-Zone-Images: true header (Jev 1.13 and 1.14) |
| Access | TypeSafe API, OpenRouter, Vercel AI Gateway |
Primary source: TypeSafe documentation.
What is a System One model?
The name comes from Daniel Kahneman's Thinking, Fast and Slow: System 1 thinking is fast and intuitive, System 2 slower and more deliberate. Large language models are built to write text for people to read. When software needs a judgment it can branch on, such as "which team should handle this ticket?", an LLM has to generate text that your code then parses.
A System One model skips the text. You give it the state and a set of questions, and it returns typed values and probability distributions in one pass. TypeSafe describes each question as a gut-check: the kind of judgment an expert could make in a few seconds with the right context. Longer reasoning should be split into several atomic questions and combined in your own code.
Coverage of Jev spread beyond developer circles through 2026: Bloomberg profiled the model in Jev, an AI Model That Can't Chat, Takes On Bigger Rivals, and enterprise-architecture writers picked up the System One category too, including Kai Waehner's How System One Models Like Jev Change Enterprise AI Architecture.
How Jev works: state, questions and answers
Every Jev request has two parts:
- State: the text Jev evaluates, such as a support ticket, an email, a document or a JSON record.
- Questions: named, typed questions about that state. You can mix types and send many in one request; each is evaluated in parallel and independently, so adding questions barely changes the response time.
Jev has four question types, which TypeSafe calls primitives. Box arrived with Jev 1.14 and works on images only:
| Question type | Use it to | Jev returns |
|---|---|---|
| Choice | Pick one option from a list (up to 255 options) | choice, a probability for each option, confidence |
| Score | Rate the state on an ordered scale (2 to 10 levels) | score, a probability for each level, confidence |
| Noul | Check whether a statement is true | noul, a probability from 0 to 1 |
| Box | Find where something is in an image | The four edges of a box, as fractions of the image's width and height |
TypeSafe trains Jev with RLCD, reinforcement learning for calibrated decisions, so its probabilities are meant to match how often it is right.
Confidence is separate from probability. It tells your code whether to act on the answer at all, for example to automate a decision only above a threshold and send the rest to a person.
Can Jev read images?
Yes. Since October 2026 an image can go anywhere text goes in a Jev request: in the state, in a question's instructions, and in its criteria. Every question type accepts images, and the new Box question type answers where something is in an image. TypeSafe labels images and Box "Dreamer-ready": good for prototypes, not yet for production, and the format, limits and accuracy can change between releases (images docs).
What a request with images needs:
- The opt-in header
TypeSafe-Danger-Zone-Images: trueon every request that contains an image or a Box question. - An
assetsmap. Each image is listed once under a reference string you choose, then written wherever it belongs. A fresh UUID per image is safest, because references are matched as plain text anywhere in the request. - PNG or JPG, base64 in a JSON body (
data_base64plusformat) or as a file part in a multipart request. Each side must be 280 to 1568 pixels; an image outside the limits is rejected, not resized. - A model version. Jev 1.13 and 1.14 both accept images. Box needs Jev 1.14.
A minimal request that asks one yes/no question about a photo and finds the dog in it:
curl https://api.typesafe.ai/v1/systemone \
-H "Authorization: Bearer $TYPESAFE_API_KEY" \
-H "TypeSafe-Danger-Zone-Images: true" \
-H "Content-Type: application/json" \
-d '{
"model": "jev-latest",
"state": {"photo": "a3f1c2e8-found"},
"assets": {"a3f1c2e8-found": {"type": "image", "format": "jpg", "data_base64": "/9j/4AAQ..."}},
"questions": {
"has_collar": {"type": "noul", "instructions": "Is the dog in `photo` wearing a collar?"},
"dog": {"type": "box", "instructions": "the dog in `photo`"}
}
}'
A Box answer gives the four edges of the smallest rectangle around the target, top, bottom, left and right, each a fraction from 0 to 1 measured from the top-left corner, plus a confidence. Multiply by the image's width and height to draw or crop the box. Box takes no criteria, and the instructions can simply name the target ("the dog in photo").
Laya reads text only. For open-weight decision models that take images, see Cloudflare Clef.
How much does Jev cost?
According to TypeSafe's models page, Jev 1.13 costs $0.042 per million input tokens ($42 per billion), and output tokens are free. Because many questions share one state, sending ten questions in one request is much cheaper than ten separate calls.
The same price applies through Vercel AI Gateway. Rate limits are currently 250,000 tokens per second and 1,200 requests per minute, and TypeSafe says they may change without notice while demand is high. Custom and enterprise plans offer higher limits.
For comparison, open models such as Laya run on your own hardware at no per-token cost. See Jev alternatives for the options.
How to use the Jev API
1. Directly from TypeSafe
Create an API key in the TypeSafe console, then send a POST request to https://api.typesafe.ai/v1/systemone:
curl -X POST https://api.typesafe.ai/v1/systemone \
-H "Authorization: Bearer $TYPESAFE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "jev-latest",
"state": "I have been trying to connect Stripe for 3 days and it keeps failing. Please help ASAP.",
"questions": {
"department": {"type": "choice", "instructions": "Which team should handle this?",
"criteria": {"billing": "Payment or subscription issues",
"technical": "Bugs or integration problems",
"sales": "Pricing or account questions"}},
"is_urgent": {"type": "noul", "instructions": "The message conveys urgency"}
}
}'
The response lists each answer under its question name, plus token usage. Official client SDKs exist for Python (pip install typesafe-sdk) and JavaScript/TypeScript (npm install @typesafe-ai/sdk). New direct sign-ups have at times been placed on a waitlist; the two gateways below are an alternative way in.
2. Through OpenRouter
OpenRouter serves Jev on a dedicated System One endpoint, https://openrouter.ai/api/v1/systemone, with the same request body and an OpenRouter API key. Use the model ID typesafe/jev-1.13 or typesafe/jev-latest.
Do not confuse it with Jev Router (typesafe/jev-router) on OpenRouter's chat endpoint. Jev Router is a different product: it uses Jev to pick which LLM and reasoning effort should answer a chat request.
3. Through Vercel AI Gateway
Vercel AI Gateway lists Jev as typesafe-ai/jev. Its TypeSafe-compatible endpoint is https://ai-gateway.vercel.sh/typesafe/v1/systemone, called with an AI Gateway API key. The Vercel AI SDK can also call it through its evaluation API.
Using Jev from coding agents
TypeSafe publishes an agent skill for Claude Code, Codex and similar tools. It gives the coding agent the context it needs to write Jev integrations: the question types, the recommended patterns and best practices for structuring questions. See Jev and Laya in Claude Code for setup.
Projects built on Jev
The best-known example is jev-ultrafast from the browser-use team, a browser agent in which Jev picks each operation and page element, and a small LLM writes text only when something has to be typed. For open-weight models that accept the same requests, see Jev alternatives.
How accurate is Jev?
In our System One benchmark, Jev 1.13 scored 0.963 on SST-2 sentiment, 0.927 on TREC question types and 0.813 on the 77-intent Banking77 set, with the best calibration of every model we tested. On the same 150 utterances in eight languages it averaged 0.87 outside English, well ahead of the open models.
TypeSafe itself says English is where accuracy is best and that other languages, including CJK scripts, are handled but not equally well.
Jev's limits
TypeSafe documents known weak spots for Jev 1.13 on its jaggedness page:
- Literal reading. Jev answers the question as written, not what you meant. Put exact conditions and boundary cases in the instructions.
- Math and counting. Counting, numeric comparisons and date arithmetic are unreliable; do them in code.
- Large, noisy state. Accuracy drops when the state is full of irrelevant detail. Filter the input first.
- Adversarial and contradictory input. Prompt-injection style content and conflicting criteria can mislead it.
- No generation. Jev cannot write text or explain its answer.
- Images are experimental. TypeSafe labels image input and Box questions "Dreamer-ready": fine for prototypes, not yet for production, and the format and limits can change between releases (images docs). Audio and video must be converted to text first.
Is Jev open source?
No. Jev's weights are not public, and TypeSafe does not fine-tune Jev per customer: every account uses the same weights, shaped through the state and the question wording. Jev is not trained on customer requests.
If you need open weights, local deployment or fine-tuning, several open System One models accept Jev-style requests, including Laya, Kev and Von. See Jev alternatives, the System One models comparison and Laya vs Jev.
FAQ
What is Jev AI?
Jev is TypeSafe AI's hosted System One model. It answers typed questions (choose, score, true or false) about a piece of text and returns probabilities instead of generated text, so software can act on the answer directly.
Who makes Jev?
TypeSafe AI. The model is available through TypeSafe's own API and through OpenRouter and Vercel AI Gateway.
How much does the Jev API cost?
$0.042 per million input tokens. Output tokens are free.
Can I use Jev with OpenRouter?
Yes. Call https://openrouter.ai/api/v1/systemone with the model typesafe/jev-1.13 or typesafe/jev-latest and an OpenRouter key.
Jev vs LLM: can Jev replace an LLM?
For narrow, structured decisions such as routing, classification, moderation and triage, it can replace an LLM call and costs far less. The difference is the output: an LLM generates text token by token that your code has to parse, while Jev returns a typed answer with probabilities for every option in one pass, and output tokens are free. It cannot write text, reason step by step or answer open questions. For the same comparison with an open model, see Laya vs LLMs.
Is there a free or local version of Jev?
Jev itself runs only as a hosted API. Open models such as Laya, Kev and Von offer local, Jev-compatible alternatives. You can try Laya in the browser on our Playground.
Related
- Jev alternatives: open System One models you can self-host
- Laya vs Jev: open weights, speed, accuracy and tradeoffs
- System One benchmark: Jev, Kev, Laya, Von and GLiNER on the same questions
- What is Laya?: the open-source decision model
OpenAI launched a competing hosted endpoint on September 29, 2026: see OpenAI Decisions API vs Jev and Laya.
Last verified against TypeSafe's documentation: November 4, 2026.