What is Jev?It is not a chatbot. Here is the difference, explained like you are five.

This week a new AI model called Jev showed up everywhere. People posted numbers that look made up: nearly two thousand ads sorted in nineteen seconds for twelve cents. A whole website audited for twenty one cents while Claude Opus was still on page 21. A browser that clicks before you finish your sentence.

And under every one of those posts, the same reply: "OK but what IS it? Is it like ChatGPT?"

No. It is not a better ChatGPT, a smaller ChatGPT, or a cheaper ChatGPT. It is a different kind of machine, and once you see the difference the numbers stop looking fake.

Imagine two friends.

The first friend talks. Ask anything, and they answer in sentences. They write your birthday poem, your cover letter, your Python script. They produce the answer one word at a time, left to right, and you wait while the words come out. That is ChatGPT. That is Claude. That is every large language model, or LLM. They are talkers.

The second friend never talks. Not one word. But show them something and give them a few buttons, and they will press the right one, instantly, and tell you how sure they are. Show them a customer email and three buttons, "Refund", "Question", "Complaint", and they press "Refund. 94 percent sure." Show them a web page and ask "should this link to that one, yes or no?" and they answer "no, 88 percent." That is Jev. It is a pointer.

That is the whole difference. A talker makes words. A pointer makes decisions.

Jev cannot answer an open question at all. It only answers three shapes of question, and you have to give it the shape in advance:

1. Which one? Pick from a list you provide.
2. How much? Rate something on a scale you define.
3. Yes or no? Give the probability that a statement is true.

Every answer comes back with a confidence number. And that number is honest, which matters more than it sounds. Chatbots are famously overconfident; ask one for a probability and it will make one up. Jev was trained with a method its makers call Reinforcement Learning for Calibrated Decisions, so when it says 90 percent it is right about 90 percent of the time. That lets software do something simple and powerful: trust the answer when confidence is high, and hand the case to a human, or to a slower model, when it is low.

There is one more consequence of the three-question rule. Because the possible answers are fixed before Jev runs, it cannot invent one that is not on the list. It cannot hallucinate a fourth button.

Here is the part most explainers get wrong.

A talker writes one word, then the next, then the next. Each word is a separate step, and every step costs time and money. That is why an LLM's output words cost several times more than its input words, and why a long answer takes seconds.

A pointer does not write anything. It reads what you gave it and produces all of its answers in one shot, at once, in a single pass. There is no stream of words to wait for. Jev's makers, TypeSafe AI, say a call takes between 70 milliseconds and half a second. Output is free; they do not even meter it. Input costs about four cents per million tokens, which works out to roughly four ten-thousandths of a dollar per decision.

That is the entire reason the posts look like magic. Sorting 1,891 ads is 1,891 "which one?" questions. Auditing 586 pages for internal links is 8,790 "yes or no?" questions. A talker pays for words it does not need. A pointer pays for nothing but the reading.

These are the public examples from this week, in the words of the people who ran them.

These are vendor and user numbers, not independent benchmarks. On TypeSafe's own four-workflow test, Jev matched a mid-tier frontier model on accuracy and sat a few points below the best ones. It trades a little peak accuracy for a lot of speed and cost. That is the honest shape of it.

Jev cannot write your email. It cannot write code. It cannot summarize a document, explain its reasoning, or hold a conversation. It gives you a number, not a rationale. If you need words, you still need a talker.

So it is not a replacement for ChatGPT or Claude. It is a different tool for a different job, and a lot of software will end up using both: the talker to write, the pointer to decide.

Here is the thing to remember. Most of what we ask chatbots to do inside real products is not writing at all. It is "which queue does this ticket go to?", "is this review positive?", "is this lead worth calling?", "which of these five is the best match?". Those are "which one?" questions in disguise, and we have been paying talker prices, and waiting talker seconds, to get pointer answers.

Jev is a bet that this is most of the work. Its name comes from Jevons, the economist who noticed that when steam engines got more efficient, people did not use less coal. They used far more. Make decisions this cheap, and software starts asking a million tiny questions it never asked before.

Talker for words. Pointer for choices. That is the whole explanation.