What is Jev, and why it is not a chat
TypeSafe's Jev returns typed choices, scores and yes/no probabilities instead of drafting text. We examine the official API and limits.
Product claims checked against the official documentation linked below. Vendor benchmarks are attributed to the vendor.
Jev is a model from TypeSafe AI, announced on 15 September 2026. The pitch is short: unstructured state in, a typed decision out. No paragraph. No JSON you then have to parse because the model decided to wrap it in a sentence.
That is the whole difference with a chat. A chat is asked to talk, and you hope the talk is shaped like data. Jev is asked a question whose answers you already listed. It picks, and it attaches a probability. If you wanted a paragraph, you called the wrong thing.
The request is a form, not a prompt
Official endpoint: POST https://api.typesafe.ai/v1/systemone. Current model id on the docs page: jev-1.13.0. Aliases such as jev-latest exist; pin the version in anything you would leave running.
A call has two parts.
The state is your data: a string, a JSON object, or an array of text values. Text only. No image, audio or video. If the thing you care about is a screenshot, you describe it first, or you do not send it.
The questions are typed. Three kinds showed up in the first week of tutorials and match the shape of the API:
| Type | You define | You get back |
|---|---|---|
| Choice | A closed list of labels, each with a criterion | One label, with a probability |
| Score | An ordered scale (calm → angry, 0 → 5) | A point on that scale, with a probability |
| Noul | A yes/no statement | A probability it holds |
“Noul” is their name for the boolean. It is not a fourth product.
A support desk is the clean example. State: the customer message. Questions: which queue, how urgent, whether they asked for money back. Your code reads answers.queue and routes. A separate model, if you still want one, writes the reply. Jev never had that job.
The TypeScript client is @typesafe-ai/sdk. The Python package is typesafe-sdk. Vercel AI Gateway, OpenRouter and Cloudflare AI Gateway also exposed it in the first days, under their own model ids. Same idea, different bill.
What “it cannot hallucinate a type” does and does not mean
The useful structural claim: the answer is one of the options you declared. You do not get a string that fails JSON.parse, and you do not get a department called "maybe billing?". That class of bug is gone.
The claim people then make, and that the docs do not support: that the choice is correct. Jev can return a valid, confident, wrong label. A probability of 0.91 is not a fact about the customer. It is the model’s score on the question you wrote. A vague criterion produces a vague decision, still typed.
There is no rationale field. If you needed the reason, you either log the state yourself or you ask a chat model afterwards. Adding “explain your choice” as a question does not turn Jev into a writer. It is not that kind of model.
Price and limits, from the model card
Checked against docs.typesafe.ai/models on 2 October 2026.
| Official card | |
|---|---|
| Input | $0.042 per million tokens ($42 per billion) |
| Output | not charged |
| Context | 64k tokens per request; 32k for the state plus the longest question |
| Choice size | tutorials cite a cap around 255 options |
| Input | text only |
| Self-host | no model file. Inference stays on TypeSafe or a gateway |
Rate limits on that page were moving the same week (“adjusting dynamically”, their words). Do not hardcode the number from a blog post, including this one.
The arithmetic is why it spread. A frontier chat at $2–$10 per million input tokens is the wrong meter for “which of these four queues?” asked on every row. At $0.042, a million of those classifications is a coffee. That comparison dies if your gateway marks the call up, or if the state you send is a full page of HTML on every keystroke.
The sites that are not TypeSafe
Search for Jev and you land on hosts selling prepaid keys (jv_live_…) against their own /api/v1/decide. Packs seen the same week were $0.25 to $0.42 per million input tokens, several times the official card. They are a reseller or a wrapper. The official host is api.typesafe.ai, the official docs are docs.typesafe.ai.
If a tutorial’s price does not match $0.042, check the host before you quote it.
The demo that made it a programming story
browser-use/jev-ultrafast is the build that left the ML thread. It is a browser agent, MIT, from the Browser Use people. Each step, the page becomes a numbered table of elements. Jev picks an operation (CLICK, TYPE_TEXT, SCROLL_DOWN, DONE, …) and a target index. A small chat model writes text only when the operation is TYPE_TEXT.
Their own measurement file: a Zürich to London search on Google Flights in 7.1 seconds, and a drop in browser protocol calls from a median of 1,092 to 101 against their previous loop. Those are the project’s numbers, on their README, not a bench I ran. The point for a programmer is the split. The decision is a function call. The prose is a different call, used rarely.
That is also the pattern in the smaller tools that showed up the same week: Jev picks the tool, Codex or Claude writes the code. The router is not the author.
Where I would not use it
- Anywhere the output is the product. A post, an email, a patch. Wrong model.
- On a choice list you cannot enumerate. “What should we do?” is not a Choice.
- As a source of truth. The probability is a threshold for your code, not a citation.
- Behind a wrapper whose price you have not read.
A reasonable gate: above a probability you chose on your own labelled set, the code acts; below it, a person or a slower model sees the case. Both paths written in advance. That threshold is yours. Copying 0.8 from a launch thread is how you auto-route the angry customer to sales.
Frequently asked questions
Is Jev an LLM? It is a model. It is not a chat model. It does not generate the string you show a user.
Can I run it locally? No file to download. Calls go to TypeSafe or to a gateway in front of it.
Why is one tutorial 10× more expensive? Different host. Official card is $0.042 per million input tokens. Several prepaid wrappers are not.
Does a high probability mean it is right? It means the model is sure about the option it picked. Sure and right are different. Calibrate on your data.
Takeaway
Jev is a typed decision call: state in, one of your labels out, with a probability, at a price that makes per-row classification boring. It is not a chat, it will not explain itself, and a valid answer can still be the wrong one. The programming move is to let it route, and to let something else write.