What Is Jev? The Fast AI Model That Answers in Types

A chat model reads your prompt and writes a reply. Jev reads the same kind of plain language and hands back a decision in a fixed shape. The System One page is blunt about the trade: “System One models do not write replies, produce code, or generate explanations of their reasoning.” You get the answer itself and the probabilities behind it, which is what a script needs to pick a path.

Those probabilities are the second thing worth knowing. TypeSafe says its System One models are “trained for calibrated decisions: their probabilities are optimized against outcomes to reflect uncertainty.” A 0.94 and a 0.55 are meant to carry different information, so your code can treat them differently. Every model TypeSafe lists is served by the same endpoint, POST /v1/systemone, per the model page.

Jev's whole vocabulary is three answer types, and you pick one per question. The primitives page sets out what each one returns, and the important line for a business owner is that “every answer is constrained to the options you supplied.” The model can't wander off and invent a category you never wrote down.

You can ask several questions in one request, and each answer comes back under the key you chose for it. That matters for cost and for speed, since one call over a page of text can settle a dozen separate judgments about it.

Here's the shape, using a form submission from a roofing site. The message goes in as the state, you name the question whatever you like, and the answer arrives under that same name.

{
  "state": "Roof leaking over the kitchen since last night's storm. Can someone come today?",
  "model": "jev-latest",
  "questions": {
    "is_emergency": {
      "type": "noul",
      "instructions": "Does this message describe active damage that needs a visit today?"
    }
  }
}

The reply carries the model id, your answers keyed by the names you picked, and a usage block with input and output token counts. For a Noul question the answer is a single number between 0 and 1, so a value near 1 tells your code to push the message in front of a dispatcher now and a value near 0.5 tells it to let a person read the thing. Any number you see in an example is an illustration of the shape, since the real value depends on the message. The request format, the response format and the retry advice all live on the API page, which says that a 429 response means you should retry with exponential backoff.

The model page prints the price as $42 per Btok and $0.042 per Mtok, and explains that “A Btok is a billion tokens and an Mtok is a million tokens.” It also says “Output tokens are free,” so you're charged for what you send and the answer costs nothing on top. Context is capped at “64k tokens per request; 32k tokens for state plus the longest question.” Input is “Text only. String, JSON object, or array of text values. No image, audio, or video input.”

Throughput is published as “250,000 tokens per second / 1,200 requests per minute,” with TypeSafe adding that the limits “can change without notice” while it serves current demand, and that higher limits come with custom and enterprise plans. A request over either limit returns a 429, which is the signal to slow your batch down.

Here's the arithmetic on a batch, and every token count in it is my assumption. Say you push 1,000 inquiries through one Noul question each, and say each request works out to roughly 1,200 input tokens once you count the message plus the instructions. That's 1.2 million input tokens. At the published $0.042 per million, the batch costs about five cents, and the output side adds nothing. Before you plan a budget around a figure like that, run fifty items and read the real usage counts in the responses.

Jev earns its keep where you have a pile of items and one repeatable judgment to make about each one. Four patterns cover most of what a marketing team or an owner actually runs into.

• Sorting. Six months of form submissions dropped into service categories, using a Choice question with one option per service line and a short description of what each one covers.
• Scoring. Every page on your site rated against a scale you wrote yourself, such as how clearly the page names the service and the town. Once the whole site carries a number, picking which fixes come first stops being a debate.
• Routing. New inquiries sent to the right inbox or crew. That's the same call a coordinator makes in the first ten seconds of reading, and it fits the handoff rules in lead triage automation.
• Checking. Whether a batch of pages, listings or ad descriptions each contain the thing you said they'd contain, run as a Noul per rule.

The common thread is volume plus a decision you could explain to a new hire in two sentences. If you've been putting off a cleanup because it means reading nine hundred things, that's the shape of job this model was built for. It also pairs with the operational work covered in AI automation for small business operations.

TypeSafe publishes a page on where this version of the model falls down, which is more useful to an owner than any benchmark would be. Five items are worth reading before you wire anything into a live process, and they're all on the jaggedness page for version 1.13.

• It takes your wording literally. The model answers the exact question you typed, so the documented fix is to “state the exact condition in the instructions. Be specific. Put boundary cases in the criteria.”
• Math belongs in your code. “Jev is not a calculator,” the page says, and the advice is to “implement any mathematical logic in code,” because the model “will perform better on semantic questions than mathematical ones.”
• Dates read as text. Version 1.13 treats a date as a string of characters, so the documented split is to let the model pull the date out of the text and keep every comparison and subtraction in your code.
• Noisy state costs accuracy. “Accuracy falls as the state grows with content unrelated to the decision,” so the advice is to “retrieve and filter in code first, and send only the fields the question needs.” Dumping a whole page of menus and footer text into a question about pricing makes the answer worse.
• It won't write for you. The page describes the model as “not trained to generate text.” Where the answer space is bounded, the documented move is to turn the extraction into a Choice over the options you already know.

Say Jev scores your 600 pages and 84 of them come back weak on clarity. Writing the replacement copy for those 84 pages is a chat model's job, or a copywriter's, and the same goes for the reply to an unhappy customer, the crawl script that gathered the pages in the first place, and the explanation of why the scores landed where they did. Jev gives you the list and the ranking, and the words come from somewhere else.

The confidence page gives a sensible way to divide the work. A high confidence answer can “act automatically,” a middling one should “proceed with caution,” and a low confidence one goes to a person or a fallback. TypeSafe puts the rule plainly: “Different actions within the same system should be gated at different levels depending on the consequences of getting it wrong.” The page also advises that you “start with conservative thresholds, test with your own data, and adjust as you observe results.” Deleting a page and flagging a page for review deserve very different bars. We keep a person on the last step of anything a customer reads, which is the same principle laid out in human review for AI marketing.

You don't need a project plan for this. Pick one decision you already make by hand fifty times a week, such as which inquiries are emergencies or which pages fail to name a town. Write the question the way you'd explain it to a new hire, including the edge cases that trip people up. Then label fifty items yourself, run the same fifty through Jev, and look at where the two disagree, because that gap tells you whether your wording or your expectations need the work.

From there, set the threshold where a wrong answer costs the least and send everything under it to a person. Keep the pile small at first, read the usage counts, and only widen the batch once the disagreements stop surprising you. The same discipline applies to any AI step you add to a marketing process, which is the thread running through what digital marketing covers today.

If you'd like the page by page view of your own site before you write a line of code, the free website audit hands you a scored list you can work down in the same order.