Most of the AI inside a business automation isn’t writing anything. It’s deciding: is this email a complaint or a question? Is this message a real customer or a sales pitch? Can this reply go out on its own, or should a person check it? Until now we used chat models such as ChatGPT or Claude for those decisions, and paid for them to write text we threw away. Jev, from TypeSafe, is a model built only for that kind of decision. Here’s what it is, how it works, what it costs (we measured it on our own website) and where it does and doesn’t fit.
The short answer
- What it is: Jev is an AI model from TypeSafe, a San Francisco startup, launched in September 2026. It doesn’t write text. You give it some information and a question with the possible answers, and it returns the answer with a probability.
- What it’s for: classifying, routing and checking. For example, sorting messages by type, deciding whether an AI-drafted reply can be sent without review, or telling a real lead from spam.
- What it costs: $0.042 per million input tokens, and the answers are free. Classifying one message from our contact form costs $0.00003: about 3 cents per thousand messages.
- How fast it is: in our tests, under a second per message, including the round trip from Spain.
- What it isn’t: it can’t write replies, explain its answers, do maths or compare dates. It works alongside a chat model and your own code, not instead of them.
- How we use it: in production on our own website, to separate real enquiries from sales pitches and junk in the contact form. On a client’s platform, to spot users swapping phone numbers or social media handles in the chat to close deals outside it. And in the automations we build for clients, to decide what goes out automatically and what a person reviews.
What is Jev?
Jev is the first model in what TypeSafe calls System One models: AI models that make fast, structured decisions for software to use directly. The name comes from Daniel Kahneman’s Thinking, Fast and Slow, where “System 1” is the fast, intuitive way of thinking we use for judgments that take a few seconds. Jev itself is named after the economist William Stanley Jevons, whose paradox says that when something becomes much more efficient, we end up using far more of it. That’s TypeSafe’s bet for AI: if decisions become almost free, they’ll be used everywhere.
TypeSafe was founded in 2024 by Diogo Almeida, Erik Gafni and Sasha Sheng. Almeida worked at OpenAI on the research that led to ChatGPT. The company spent two years in stealth and launched Jev in early access on 15 September 2026, alongside a $40 million seed round. Three weeks later, in October, it raised an $870 million Series A led by Andreessen Horowitz, at a $7.5 billion valuation. TypeSafe says a third of Fortune 500 companies already use Jev, although it hasn’t named them. Version 1.13 arrived on OpenRouter on 18 September, and version 1.14, which also accepts images, as a preview on 14 October.
The difference from ChatGPT, Claude or Gemini is in what comes out. A chat model generates text, word by word, and if you want a decision you have to ask for it in a particular format and then check the format came back right. Jev never generates text. You define the possible answers in advance, and it returns a probability for each one.
How Jev works: information in, decisions out
Every call to Jev has two parts:
- The state: the information to judge. A customer’s message, an email, an order with its history, a contract clause. It can be plain text or a structured object with fields.
- The questions: what you want to know about that state, each with its possible answers and a short description of when each applies.
There are three types of question:
| Type | What it asks | What it returns | Business example |
|---|---|---|---|
| Choice | Which of these options fits? | The most likely option, a probability for each, and a confidence level | “Is this ticket about billing, delivery, returns or something else?” |
| Score | Where does this fall on a scale? | A position on the scale and a probability for each level | “How upset is this customer, from calm to furious?” |
| Noul | Is this true or false? | The probability that it’s true | “Does this message ask for a refund?” |
Version 1.14 adds a fourth type, Box, which finds where something is in an image. TypeSafe describes it as an early feature.
All the questions in a call are answered at the same time and independently, so asking five questions costs barely more time than asking one. And because the answer can only be one of the options you defined, Jev can’t reply with something your software doesn’t expect, such as a made-up category or a broken format.
What it returns isn’t just “yes” or “no”, but a probability. TypeSafe trains Jev so that those probabilities are calibrated: across many decisions, when it says 90%, it should be right about 90% of the time. That lets you set thresholds. Above 95%, the system acts on its own; between 60% and 95%, a person reviews it; below that, it’s discarded. TypeSafe itself points out that calibration is measured across groups of decisions and doesn’t guarantee any individual answer is right.
Jev vs ChatGPT, Claude or Gemini
| Jev | Chat models (ChatGPT, Claude, Gemini) | |
|---|---|---|
| What it returns | One of the answers you defined, with a probability | Text, which may or may not follow the format you asked for |
| Explanations | None. It doesn’t say why | It can explain its reasoning |
| Writing replies, summaries, code | No | Yes |
| Price per million input tokens | $0.042 | From about $0.20 to $10, depending on the model |
| Price of the output | Free | Usually several times the input price |
| Speed | Fractions of a second | From seconds to minutes with reasoning |
| Confidence in the answer | Always, as a probability | Only if you ask, and it tends to be overconfident |
| Best for | Classifying, routing, checking, filtering, at high volume | Writing, reasoning, conversations, anything open-ended |
TypeSafe says Jev “can’t hallucinate”. That’s true in a narrow sense: it can’t invent an answer outside your options. But it can choose the wrong option, just as a person can. The real advantage is that it tells you how sure it is, so the doubtful cases can go to a person instead of slipping through.
In practice, the two work together. A typical customer-service automation uses a chat model to draft the reply and Jev to decide whether that draft can be sent without review. We explain that model in detail in our guide to AI customer service for ecommerce.
How much does Jev cost?
Jev charges only for what you send it: $0.042 per million input tokens (TypeSafe also quotes it as $42 per billion). Output tokens are free. On OpenRouter, the platform we use to call it, there’s no subscription or minimum.
To give a real figure rather than a theoretical one, we measured our own contact-form check. Each message, including the description of our company and the two questions, is about 715 tokens:
| Volume | What it costs with Jev |
|---|---|
| 1 message | $0.00003 |
| 1,000 messages | About $0.03 |
| 100,000 messages | About $3 |
| 1 million messages | About $30 |
For comparison, doing the same check with a small general-purpose model such as Claude Haiku 4.5 ($1 per million input tokens and $5 per million output) would cost just under a tenth of a cent per message, nearly 30 times more. With a frontier model, the difference runs into the hundreds. TypeSafe’s own benchmark claims Jev is up to 444 times cheaper and 193 times faster, but those figures come from workflows designed by its own team, and TypeSafe itself says they’re at the high end of what to expect.
Two caveats. TypeSafe says openly that it can’t prove its prices aren’t subsidised, although it expects them to fall rather than rise. And the endpoint used to call Jev on OpenRouter is still marked as alpha. For a small business, though, the cost of the model is no longer a factor in the decision: what costs money is designing the questions well and connecting them to your systems.
How we use Jev: telling real leads from spam
Our contact form gets the same thing every company’s does: SEO agencies promising first place on Google, offshore development teams, lead lists, funding offers and a fair amount of automated junk. Mixed in with all that are the messages that matter: a clinic that wants its WhatsApp to answer patients, a manager asking for a quote.
Classic filters don’t solve it. Keyword lists block real customers who mention “SEO” and let through pitches written slightly differently. A captcha stops bots, but not a person pasting a template. So the form has several layers: a hidden field that only bots fill in, a check on how long the form took to complete, a limit on messages per connection and, finally, Jev.
We ask Jev two questions about each message. The first is a Choice with four possible answers:
"kind": {
"type": "choice",
"instructions": "What kind of message is this submission to the website's contact form?",
"criteria": {
"lead": "A potential client asking about AITOMATIK's services, even if short, vague or badly written, in any language.",
"other_legit": "A real person with a legitimate reason: job application, existing client, a partnership idea specific to AITOMATIK.",
"sales_pitch": "Someone selling their own services to AITOMATIK: SEO, link building, marketing, lead lists, outsourced development...",
"junk": "Scam, phishing, crypto, adult content, gibberish or automated junk."
}
}The second is a Noul: “Is this submission spam that the team shouldn’t be notified about?”. Along with the message, we send Jev a short description of who we are and what we do, so it can tell an enquiry about our services from a template that was never written for us.
The rules that act on the answer live in our code, not in the model:
- If the probability of “sales pitch” plus “junk” reaches 80%, the team doesn’t get a notification.
- Nothing is deleted. Every message is stored and can be checked in our admin panel, with the probabilities Jev gave it.
- If the check fails (a timeout, an error), the message counts as genuine. A real lead must never be lost because a filter failed.
To show what the answers look like, we ran three typical messages through it while writing this guide. A clinic in Palma asking, in Spanish, for WhatsApp to answer its patients: 100% lead. An English SEO pitch promising first place on Google: 96% spam. A Spanish message from an offshore development agency offering developers at $15 an hour: 88% spam. Each answer took under a second and cost three thousandths of a cent.
Another example: keeping deals on the platform
On marketplaces and platforms where clients and providers talk through a chat, there’s a well-known problem: once they’ve met, some users swap phone numbers and close the deal outside the platform, and the platform loses its commission. We built a check for exactly this on a client’s platform.
Spotting a full phone number or an email address is easy with code: a regular expression catches them. The hard part is everything else people do when they know they’re being watched:
- “Find me on Insta as laura.design”.
- “My number starts with six one two…” and the rest in the next message.
- One digit per message, spread over a dozen messages.
- “Add me on the green app”, “search for my name on LinkedIn”, “I’ll send you my details in my portfolio”.
A filter based only on code either misses these or blocks half the normal conversation. So the check combines both: code catches the obvious patterns, and Jev reads the last few messages of the conversation together and answers a Noul question: “Is someone trying to share contact details or move the conversation off the platform?”. With the conversation as context, it catches a number split across several messages or a handle written to slip past the filter. Depending on the probability, the platform shows a warning, hides the message or flags the conversation for the team to review.
The reason it’s viable is the price. Checking every message on a busy platform with a chat model would cost more than many platforms earn in commission. With Jev, it’s a few cents per thousand messages. And it isn’t the only option: other decision models do something similar, so it’s worth testing them on your real conversations. If you do this, tell your users in your terms that the chat is monitored for this purpose: it’s personal data, and GDPR requires you to inform them.
What else Jev is good for in a business
The pattern is always the same: information comes in, a decision comes out with its probability, and the rules that act on it live in your code. TypeSafe’s own list of use cases and the projects published so far point to these as the most useful for a business:
| Area | Decisions Jev can make | Example |
|---|---|---|
| Customer service | Type of query, urgency, frustration, risk of leaving, refund requests, whether a reply follows your policy | Sending each ticket to the right queue, and letting only the safe replies go out on their own |
| Sales and leads | Fit with your ideal customer, buying intent, which service they’re interested in | Prioritising inbound leads before anyone on the team reads them |
| Recruitment | Relevant experience, required skills, fit with the role | Screening thousands of applications and sending the doubtful ones to a recruiter |
| Marketplaces and platforms | Prohibited or counterfeit listings, fake reviews, contact details shared in the chat | Reviewing every listing and message without a team reading them all |
| Moderation | Insults, harassment, spam, fraud, personal data exposed, unsubscribe requests | Comments, reviews and community or WhatsApp group messages |
| Insurance and claims | Complexity, missing information, signs of fraud | Separating the simple claims that can be processed straight away from those that need an adjuster |
| Compliance and legal | Missing clauses, advertising claims you can’t make, policy breaches | Checking contracts or marketing copy before they go out, and escalating the doubtful ones |
| Documents and email | Document type, department, deadlines | Invoices to finance, CVs to recruitment, notices with a deadline to the person in charge |
| Checking other AIs | Whether a draft answers what was asked, whether it cites the source correctly, whether it promises something it shouldn’t | The “can this go out on its own?” check before an AI reply is sent |
| Choosing the model | How complex a request is | Sending simple questions to a cheap model and complex ones to a powerful model or a person |
The best-documented example so far is TypeSafe’s case study with Jack & Jill, a recruitment marketplace that uses AI to match candidates with companies. It replaced Gemini 3.1 Flash Lite with Jev in one stage of its candidate screening: the cost per thousand candidates fell from $0.755 to $0.092 (88% less), the median time halved from 20.3 to 10.3 seconds, and the quality of the shortlist didn’t change in any statistically significant way. It’s a case study published by TypeSafe itself, but the figures are specific and include the baseline.
We already describe this approach in our pages on AI for clinics, restaurants, accounting firms and property management: the AI drafts, Jev decides what can go out on its own, and a person handles the rest.
Limits and risks of Jev
Jev is new, and TypeSafe itself publishes a list of its weak points. These are the ones that matter most for a business:
- English is its strongest language. TypeSafe says other languages are supported but not equally well, and recommends testing with your own content. Our Spanish tests went well, but test your real messages before relying on it. A trick that works for us: write the questions and criteria in English, even when the messages arrive in another language.
- It reads literally. It answers exactly what you wrote, not what you meant. If a category is ambiguous, it’ll get it wrong. Writing clear criteria for each option is most of the work.
- No maths, dates or counting. Whether an invoice is over €500 or a date falls within the return window should be calculated in code, not asked of the model.
- Content written to fool it can sway it. A spam message that insists it’s a legitimate enquiry can shift the probability. That’s why we combine it with other filters and never delete anything automatically.
- Option order matters a little. TypeSafe has seen it lean slightly towards the first option in a Choice. When testing, change the order and check the answer holds.
- Long texts lower the accuracy. It works best with just the information the question needs, not a whole document full of irrelevant detail.
- Personal data. Requests go through OpenRouter to TypeSafe in the United States. TypeSafe says it doesn’t train Jev on customer data, and offers a data processing agreement and zero data retention for enterprise customers. If you send it names, emails or customer messages, include it in your GDPR records and send only what the decision needs. We explain how to approach this in our guide to AI and GDPR.
- It’s early. The OpenRouter endpoint is alpha, versions change every few weeks, and most of the performance figures are TypeSafe’s own. Fix the model version you’ve tested (
jev-1.13,jev-1.14) instead of always using the latest.
How to start with Jev in your business
- Pick one repetitive, bounded decision that’s made many times a day: sorting messages, approving replies, filtering contacts.
- Collect a hundred real examples with the answer a person would have given.
- Write the questions and the criteria for each option, clearly and specifically.
- Run it in shadow mode: Jev decides, but a person still does the work, and you compare.
- Set the thresholds from what you’ve seen: when it acts on its own, when it asks for review.
- Turn it on gradually and keep reviewing the doubtful cases.
How we do it at Aitomatik
We use Jev in our own website and in the automations we build for clients, as the layer that decides what can be done automatically and what needs a person. We design the questions with your team, test them on your real data before switching anything on, and connect them to your systems, inside the automation itself. If it turns out a simple rule is enough, we’ll tell you.
Frequently asked questions
Is Jev an LLM? Not in the usual sense. It understands natural language like a large language model, but it doesn’t generate text: it only returns probabilities for the answers you define. TypeSafe calls it a System One model.
Can Jev hallucinate? It can’t invent an answer outside the options you give it, or return a broken format. It can choose the wrong option, so you should set thresholds and send the uncertain cases to a person.
How much does Jev cost? $0.042 per million input tokens, and the output is free. In our case, classifying a contact-form message costs $0.00003, about 3 cents per thousand messages.
Does Jev work in Spanish and other languages? Yes, but English is its strongest language and TypeSafe recommends testing with your own content. In our tests with Spanish messages it classified correctly, with high probabilities.
Can Jev write replies or explain its decisions? No. To write, use a chat model. Jev decides; for example, whether that model’s reply can be sent without review.
How can I access Jev? Through OpenRouter, with an API key and no waitlist, or through TypeSafe’s own API and SDKs for Python and JavaScript.
Want to automate the decisions your team makes every day?
If your team spends hours sorting messages, approving replies or filtering contacts, there’s probably a decision in there that Jev can make for a fraction of a cent. Tell us about it on a free call and we’ll show you how we’d set it up, with real figures from your case.
