What Jev is
- Who made it: TypeSafe AI, which launched Jev on 15 September 2026. The press calls its founder, Diogo Almeida, a ChatGPT co-creator. In his own words, his earlier work at OpenAI "ended up as the research behind ChatGPT".
- What goes in: a question, the facts, and the list of answers it is allowed to give.
- What comes out: one of three shapes. Choice picks one item from your list. Score rates something against a scale you set. Noul gives the chance that a statement is true, as a number from 0 to 1. Each comes with probabilities, and Choice and Score also get one confidence number. It never produces free text.
- Why anyone wants that: it is fast and cheap. TypeSafe quotes answers in 70 to 500 milliseconds, at $0.042 per million tokens of input (tokens are the word-chunks AI reads). Software can make thousands of small decisions with it, such as sorting messages or checking another AI's work.
TypeSafe's quick start (read 29 September 2026) shows all three on one sample message, from a customer who has spent three days failing to connect a payment account and is losing sales.
- Choice, which team should handle it: technical 85%, billing 15%, sales 0%.
- Score, how frustrated the customer is: "Frustrated but civil", the middle of three levels.
- Noul, whether the message is urgent: 1.0, which means yes.
Who can use it right now
Developers first, and access is still staged. The launch post says TypeSafe is "bringing developers off the waitlist as quickly as we can", and its models page (read 29 September 2026) says limits will change as it lets in more users.
Once you are in, the quick start points to a Playground where you paste some text and add a question in your browser. There is still no chat window, so you are more likely to meet Jev inside an app one day than to open it yourself.
What "can't hallucinate" means here
TypeSafe says Jev "can't hallucinate". That holds in one sense: it cannot make something up, because it can only answer from your list. It can still pick the wrong item from that list, and do it confidently.
Its selling point is that the confidence numbers are accurate. An independent check of the launch claims (read 21 September 2026) saw the list price repeated across several places, and no published paper or reliability chart yet to back the accuracy of those numbers.
The idea worth borrowing
TypeSafe's guidance to developers (read 21 September 2026) suggests sorting answers into three lanes, and setting the bar higher where a mistake costs more.
- High confidence: act on it.
- Medium: go carefully, and confirm or find out more first.
- Low: do not act, and ask a person.
You can use the same habit in Claude or ChatGPT today. Give it a fixed list, make it pick one, and decide in advance what you will do at each level of confidence.
The prompt: pick from my options and tell me how sure you are
Replace each bracketed line under THE DECISION with your answer, paste in the facts, then send.
You are helping me make one decision from a fixed list of options. I am going to act on your pick, so a confident answer built on a guess is worse for me than "I do not have enough to choose". Stay inside my list, use only the facts I give you, and be honest about how sure you are. THE DECISION (I fill this in once) - The decision, as one question: [e.g. "Which of these three quotes do I accept for the bathroom repair?", "Is this email a complaint, a question, a sales pitch or spam?", "Should I reply to this message today, this week, or not at all?"] - My options, one per line: [every answer you would accept, e.g. "Quote A", "Quote B", "Quote C". Keep it under ten. The chatbot adds "none of these" and "not enough information" if you leave them out] - What matters most, in order: [e.g. "1. finished before the 30th, 2. total cost, 3. a written guarantee"] - What happens if the pick is wrong: [e.g. "mild annoyance", "I lose a deposit of 500", "a client gets the wrong answer". Be honest, because this sets how sure the answer has to be] - The facts: [paste everything you want used: the quotes, the email, the notes. If a fact is not pasted here, it does not exist for this decision] - My rule for acting (change it if you like): High: I act on it. Medium: I check the one thing you name first. Low: I ask a person or get more facts. BEFORE YOU PICK If the decision, the options or the facts are missing or still in brackets, ask me for them and wait. If I did not say what happens if the pick is wrong, ask me once. If "what matters most" is blank or still in brackets, ask me once. If I tell you nothing ranks above the rest, or this is a sorting decision (which label fits this message), weigh each option on how well the facts match it, and in Step 4 read "the things I ranked first or second" as "the facts that separate my top two options". If two of my options overlap so that both could be right at once, say so and ask me to reword them. If my options leave out an obvious answer, tell me what it is and ask whether to add it. Add "none of these" and "not enough information" to the list if I did not. STEP 1: SAY WHAT YOU ARE WORKING FROM List the facts from what I pasted that bear on this decision, each in one line. For anything I said matters but gave no fact for, write [NOT STATED: what is missing]. Do not fill a gap from general knowledge, and do not look anything up unless I ask you to. STEP 2: WEIGH EACH OPTION For every option, including "none of these" (leave "not enough information" out, it is for Step 4), give one line for it and one line against it, using only the facts in Step 1 and my order of what matters. If an option has nothing for it, say so plainly. STEP 3: PICK ONE Name exactly one option from the list, word for word. Then give a rough split across my options and "none of these" as percentages that add up to 100, so I can see whether it was a clear winner or a close call. These are your estimates and I know they are rough. STEP 4: SAY HOW SURE YOU ARE Give one band, HIGH, MEDIUM or LOW, with the reason in one sentence. Use these rules: - HIGH only if one option is well ahead and nothing in Step 1 is marked [NOT STATED] for the things I ranked first or second. - MEDIUM if one option leads but a missing fact could change the pick. - LOW if the top two options are close, or the facts that matter most are missing. If so, change the pick to "not enough information", say which option was ahead, and leave the percentages as they were. If being wrong is costly by what I told you, move one band down from where you would otherwise land, and say that you did. STEP 5: TELL ME WHAT TO DO NEXT, BY MY RULE Apply my rule for acting to your band. For MEDIUM, name the single fact I should check first, and say which way the pick moves if that fact goes the other way. For LOW, name the two facts that would let you choose, and who or where I could get them from. RULES - Never pick something that is not on the list, and never blend two options into a new one. - Never invent a fact, a price, a date or a quote. If it is not in what I pasted, it is [NOT STATED]. - Do not soften the pick to be polite. If every option is poor, say "none of these" and why. - This is a decision aid. For anything legal, medical or financial, say that a qualified person where I live should confirm it, and carry on with the rest. CHECK YOURSELF BEFORE YOU ANSWER 1. The pick is word for word one of my options, or "not enough information" if the band is LOW. 2. The percentages add up to 100 and the band matches them: a 45 to 40 split can never be HIGH. 3. Every reason traces to a fact in Step 1. 4. Tell me the one thing most likely to make this pick wrong.
The honest bit
- A chatbot's percentages are guesses. Jev is trained to give its numbers, while a chatbot estimates them on the spot. Use the band to decide what to check.
- The guesses stay high when facts are missing. A study posted in August 2026 tested AI models on 500 medical questions, some with the right answer removed, and found that "model confidence remains misaligned with accuracy".
- This is early news. Last checked on 29 September 2026, two weeks after launch, so access, prices and the evidence may all change.
Try it on one decision today
Pick a choice you have been putting off, write the options down, and paste in the facts. If it comes back low, you have learned which facts to go and get.
Questions people ask
What is Jev for AI?
Jev is an AI model from TypeSafe AI, launched on 15 September 2026, that picks from a list of answers you set in advance and says how sure it is, instead of writing text. It is built for developers, so software can make many small, cheap decisions with it, such as sorting messages.
Can I use Jev like ChatGPT?
No. It has no chat box, and it can only answer a question where you have already supplied every possible answer. The closest you can get in a chatbot is the prompt on this page.
Does "can't hallucinate" mean it is never wrong?
No. It cannot say something that is off your list, but it can pick the wrong item from the list. Its makers tell developers to start with cautious thresholds, test with their own data, and adjust as they see results.
Why did it play Doom?
To show speed. Playing a game means deciding many times a second, which a chatbot is too slow to do. TypeSafe's launch post says its engineer ran about ten questions a second, at a cost of about $7 an hour.