What is Jev AI? TypeSafe's Jev model, running in our customer chats

Jev AI is a System One model from TypeSafe AI. It does not write text. You send it a state, such as a customer's message, with typed questions, and it answers each one with a calibrated probability: yes or no, a pick from options you listed, or a point on a scale you described. Your code acts on the answer.

I run HiQBot, an AI agent that answers customers on website chat, WhatsApp, Instagram, Messenger, Telegram and LINE. Most guides to the Jev model are written for people building coding agents, browser bots and model routers. This one covers customer conversations, from a team that tested Jev on them and then switched it on.

What is Jev AI?

Jev is the first System One model from TypeSafe AI, released in early access on September 15, 2026, the day the company came out of stealth with a $40 million seed round led by DCVC. Its founder, Diogo Almeida, was one of the primary authors of the InstructGPT paper at OpenAI, and co-founded TypeSafe with Erik Gafni and Sasha Sheng. TypeSafe trained Jev with a method it calls reinforcement learning for calibrated decisions (RLCD). Chat models are trained on the answers people prefer; RLCD rewards probabilities that match how often the model turns out to be right. The name System One comes from Daniel Kahneman's fast, intuitive thinking, and Jev from the economist William Stanley Jevons. The launch post calls it “a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”

A request carries one state and a set of named questions. Each question has a type:

Every question in a request is answered at the same time, so ten questions take about as long as one. TypeSafe publishes end to end latency of 70 to 500 milliseconds and a price of $0.042 per million input tokens, with output free because there is no generated text to meter. A request can hold 64,000 tokens, and the state plus its longest question must fit in 32,000. Input is text only, and English is the language it was mainly trained on. The current version is jev-1.13.0, and the details are in TypeSafe's model documentation.

Jev vs an LLM inside an AI agent

Jev's most useful property is the fixed shape of its answer. A text model asked which tool to call can name a tool that does not exist. Jev only ever returns one of the options you supplied. That guarantee is narrower than the launch coverage suggests, because schema safety is not correctness: Jev can still pick the wrong option, and do it with high confidence. What it cannot do is return something your code was not written to handle. In an AI agent the two kinds of model split the work like this:

Is Jev a chatbot? Can you chat with Jev?

No. Jev has no chat interface and cannot write a sentence, so you cannot hold a conversation with it the way you would with ChatGPT or Claude. People have tried. One developer built a chat-shaped command bar where Jev picks the tool and its arguments and plain code builds every reply from the tool's own data. Another turned Jev into what their Hacker News post called a “lousy chatbot”: asked to write a short story, it answered “a story”. TypeSafe's own documentation says forcing it to generate text works badly and slowly.

A customer-facing AI agent needs a model that writes. In HiQBot, Jev sits beside that model and makes the small, repeated calls around the conversation, quickly and with a number attached.

How Jev confidence works, with examples

Every Choice and Score answer comes back with a probability for each option and a single confidence value from 0 to 1. A Noul returns the probability of yes and nothing else, because that number already is its certainty. TypeSafe's confidence guide is short and worth reading in full.

A Choice: which team should handle this?

Cloudflare's Jev model page shows a real jev-1.13.0 response to the message “Help! My payouts have been failing for 3 days.” Asked which team should handle it, Jev put 0.87 on billing, 0.13 on technical and 0 on sales, and reported a confidence of 0.80.

Confidence is not the winning probability. It measures how far the leading option stands above an even split. With three options, a split of a third each scores 0, and all the weight on one option scores 1. A 0.87 lead lands at 0.80. A second strong candidate pulls confidence down fast, a sign that the message did not separate the options.

A Choice that should go to a person

OpenRouter ran 60 support tickets through Jev. One customer had paid with a gift card and a credit card, and asked where the gift card part of their refund went. Jev filed it as a return or refund at a confidence of 0.54. It was the only ticket under 0.8, and the only intent Jev got wrong. A rule that sends anything under 0.8 to a person would have caught it without anyone knowing in advance which ticket was hard.

A Score: how severe is this bug report?

TypeSafe's Score guide uses a report that an export button crashes in Safari but works in Chrome, rated on three levels: cosmetic, broken with a workaround, and blocking with no workaround. Jev put 0.57 on “workaround exists” and 0.43 on “no workaround”, which gives a score of 1.43 and a confidence of only 0.35. The model was split, and fairly so. When the middle level gained one example that looked like the report, “export fails in one browser but works in another”, the answer moved to 1.03 at 0.96 confidence. An unrelated example changed nothing. When a Score comes back with low confidence, check the level descriptions first.

A Noul: is this urgent?

On the same payouts message, a Noul asking whether it conveys urgency came back at 0.95. A Noul near 0.5 means Jev cannot tell from the message. To measure how urgent something is, ask a Score.

What calibrated means, and where it falls short

Calibrated means that across many answers given at 0.8, about 80% turn out right. It describes a group of predictions, never a single one. The most careful public check so far is Anthus's test on 8,801 labelled examples. Higher confidence did mean more often right, and every yes or no answer above 95% was correct. The raw numbers still ran high: Choice answers averaged 91.4% stated confidence against 76.1% accuracy. Fitting the numbers to a few hundred of their own labelled examples fixed most of the gap. If you run Jev on customer messages, draw your lines from your own labelled messages.

“The user has to say, okay, if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it. But if it's 95%, sure, then I can do something with it.”

The three confidence bands

TypeSafe suggests splitting confidence into three ranges, each with its own behaviour:

Where the lines sit depends on what a wrong call costs. A help article suggestion can act on a weaker answer than anything that moves money or speaks to a customer. Our own rule follows the same shape: Jev acts only where it is sure, and below that it stays out of the way.

What teams use the Jev model for

TypeSafe's use case map lists customer support next to model routing, guardrails and search: classifying tickets by issue and intent, detecting urgency, frustration, churn risk and refund requests, routing cases, and checking support responses against policy. In the first week after launch, teams published results like these:

• Vercel replaced an OpenAI model in the classifier that checks commands for safety, and got results five to 18 times faster, with better accuracy, according to TechCrunch. Vercel also put Jev on its AI Gateway.
• Browserbase uses Jev to pick the next action in its Stagehand browser framework and falls back to an LLM when Jev is not sure enough. Median time for that step dropped from 1.97 to 0.46 seconds in their early testing.
• In OpenRouter's benchmark, Jev got 59 of 60 support ticket intents right at a median of 194 milliseconds, for $0.0248 per 1,000 tickets against $2.88 for Claude Opus, and caught all 40 prompt injection tests in a second run.
• A Towards Data Science test on 77 banking intents found Jev almost twice as fast as two OpenAI models and less accurate, 79.0% against 83.9% and 86.2%. With 7 intents instead of 77 it was roughly level, and its confidence tracked its accuracy well.
• TypeSafe's own evaluation puts Jev at 67.8% agreement across four workflows against 74.1% for the best frontier model. Its strongest workflow was customer service, where Kingy AI's review reads the dashboard as 76.0% against 78.3%. The reference answers came from averaging two frontier models rather than from human review, and TypeSafe says so itself.

Every report found Jev fast. Accuracy varied with the task, the number of options and the wording of the question. None of them measured Jev on live customer conversations.

How HiQBot runs Jev in live customer chats

Jev reads every text message our AI agent answers, before the agent starts its reply. When Jev is confident, the agent gets a head start and simple questions are answered with fewer steps. When it is not, Jev adds nothing and the agent handles the message exactly as it did before. It works the same way on WhatsApp, the website widget and every other channel we connect.

• Jev starts as the message arrives, next to everything else the agent prepares, so the customer does not wait for an extra step. Its own answer takes about 0.4 seconds.
• The model that writes keeps every decision. It writes every reply and decides every booking, lead capture and handover. If Jev's call looks wrong in context, the agent sets it aside.
• Jev never acts on a customer. It does not reply, refuse, hand a conversation to a person or run an action.
• If Jev is slow or down, the message is handled exactly as it would be without Jev.
• Jev sees as little as we can give it: the business's name and industry and the customer's message, with email addresses and phone numbers masked the same way our agent sees them. No conversation history and no customer notes. TypeSafe says Jev is not trained on customer requests.
• Voice notes skip Jev, because it reads text only.
• We pin the version to jev-1.13.0 instead of the moving “latest” alias, and re-run our tests before any upgrade.
• Each message costs about a hundredth of a cent at TypeSafe's published price.

How we tested Jev before production

We did not switch Jev on until it had passed three kinds of test, run on jev-1.13.0 in the week after launch.

1. We built a set of labelled customer messages from industries including car dealerships, hotels, restaurants, travel and software. It includes messages written to trick the agent. Every message was classified twice. Every answer Jev gave with enough confidence for us to act on was correct. Where it was less sure, it added nothing.
2. We ran the same customer conversations through the whole agent twice, once with Jev and once without, and compared speed, the number of steps the agent took, and the replies themselves.
3. Our engineers also tested it by hand, outside the automated runs, before it went live.

What changed for customers:

• Short, simple questions got faster, because the agent could skip a step.
• Questions about the business's own facts, such as hours or prices, came back sooner.
• The slowest replies got quicker, and typical reply times stayed about the same.
• Long messages that ask for several things at once got slightly slower. We are still working on those.

Our review found no change in reply quality either way. Jev made simple answers faster, and that is the whole gain we claim. We set our confidence lines from the labelled messages, separately for each kind of call, by what a wrong call would cost.

Mistakes to avoid when you put Jev in front of customers

Most of these come from TypeSafe's list of jev-1.13's weak spots. Several we learned the slow way.

1. Letting Jev have the final word on a customer. In a support inbox the state is written by the person you are judging, and text written to steer the model can move its answer. We rejected any design where Jev could refuse or redirect a customer on its own. A wrong refusal loses a lead.
2. Sending the whole conversation. Accuracy falls as the state fills with detail the question does not need, and we saw the same: giving Jev more context made it slightly worse.
3. Trusting the raw confidence number. It ranks answers well, but the numbers themselves can run high. Set your lines from your own labelled messages.
4. Using Jev to grade the agent's replies. OpenRouter's check of drafts against policy looked promising on four examples. In our test, Jev marked a reply that the business's knowledge base fully supported as almost certainly unsupported, so we do not use it for that.
5. Assuming English results carry over. English is Jev's main training language. On our translated test messages it was slightly less accurate than on the English originals. Test the languages your customers write in.
6. Asking it to count or compare dates. How long since the order, whether a booking falls inside a window, how many messages remain: keep all of that in code.
7. Opening a new connection for every call. Our first version made replies slower, because setting up each connection took longer than Jev's answer. Keeping the connection open and starting Jev as the message arrives fixed it.
8. Tuning against the latest alias. The jev-latest alias moves when TypeSafe ships a new version, and your confidence lines move with it. Pin the version you tested.

Common questions

What is Jev AI?

Jev is the first System One model from TypeSafe AI, released in September 2026. You give it a state and typed questions, and it returns answers with calibrated probabilities. It does not generate text, so it cannot write a customer reply. Your code calls it the way it would call a fast classifier.

Is Jev a chatbot? Can you chat with Jev?

No. Jev has no chat interface and produces no sentences. Developers have built chat-shaped demos where Jev picks a tool and code writes the reply, but a real customer conversation needs a model that writes. Jev works beside that model, making quick decisions with a confidence attached.

Does HiQBot use Jev?

Yes. Since September 2026, Jev reads the text messages our AI agent answers. When it is confident, simple questions are answered faster. When it is not, it adds nothing. The model that writes the replies still makes every decision, and Jev never replies to or acts on a customer.

What happens when Jev is unsure or wrong?

When Jev is unsure, it adds nothing and the agent handles the message as it always has. When it is confident and wrong, the agent can still set its call aside, because Jev never acts on its own. If Jev is slow or unavailable, the message is handled exactly as it would be without it.

Does Jev see customers' personal details?

Jev sees the customer's message with email addresses and phone numbers masked, plus the business's name and industry. It does not see the conversation history or customer notes. TypeSafe says Jev is not trained on the requests customers send it.

How much does Jev cost?

TypeSafe lists Jev at $0.042 per million input tokens, with output free. For us that works out to about a hundredth of a cent per customer message. You can call it through TypeSafe's own API or through Vercel AI Gateway, Cloudflare and OpenRouter. Check the provider's page before you budget, because early-access prices can change.

What is a System One model?

It is TypeSafe's name for a model built for fast, repeated judgements rather than step-by-step reasoning, after the psychology term for intuitive thinking. In practice it means one pass that returns a typed answer and a probability, instead of generating text token by token.

Does Jev work on voice notes, images and other languages?

Jev reads text only, so a voice note has to be transcribed and an image described before it can be judged. It accepts other languages, but English is where it is most accurate. Test it on the languages your customers use. Our guide to omnichannel customer service covers running one agent across every channel.