Everyone online is talking about Jev, and almost nobody can explain it in one sentence. Here is mine: Jev is an AI model that cannot chat, write, or code. It only answers sorting questions, like which team should handle this message, and it does that in a fraction of a second for a fraction of a cent. It sounds dull. It is also why developers have talked about little else since it launched on September 15.
This post explains Jev in plain English, shows who can actually use it, and walks you through a 10 minute test you can run without writing a line of code.
What Jev Is, Without the Jargon
Jev comes from TypeSafe AI, a San Francisco startup that came out of stealth on September 15, 2026 with $40 million in seed funding led by DCVC. Its CEO, Diogo Almeida, is a former OpenAI researcher and one of the co-inventors of RLHF, the training method behind ChatGPT.
Think of the difference between an essay question and a multiple choice answer sheet. ChatGPT, Gemini, and Claude are built for essays. They write paragraphs for people to read. Jev is the person at the front desk with a clipboard. A note comes in, they tick one box, and they write down how sure they are. They never hand you a paragraph.
The company calls this a System One model, a nod to the fast, gut level thinking described in Thinking, Fast and Slow. TypeSafe’s own pitch is that Jev is a function call: messy information goes in, a typed decision with probabilities comes out, and ordinary software takes it from there.
The Three Question Types
Jev does not take open prompts. You give it some text and a set of questions, and each question is one of three types, according to TypeSafe’s documentation.
| Type | Plain English | Example question | What comes back |
|---|---|---|---|
| Choice | Pick one option from a list | Which team should handle this message? | The pick, a probability for every option, and a confidence number |
| Score | Place it on a scale | How frustrated does this customer sound? | A score, probabilities, and a confidence number |
| Noul | Yes or no | Is this person asking for a refund? | A single number from 0 to 1 |
You can ask many questions about the same message in one go, and they are answered in parallel. Here is what that looks like for a message that says my card was charged twice (the numbers below are illustrative, not a real run):
- Which team? Billing 0.91, technical 0.06, sales 0.03
- Is it urgent? 0.72
- Is the customer asking for money back? 0.88
Here’s the thing about that confidence number. It tells your software when to act alone and when to hand the message to a human. TypeSafe’s docs explain that confidence is calculated from how the probabilities are spread: one clear winner means high confidence, a flat spread means low. That one design choice is what makes Jev usable for automation.
Why Everyone Is Excited About the Speed and Price
Chat models write one word at a time. Jev returns every answer in a single pass. According to TypeSafe’s launch post, responses take 70 to 500 milliseconds. Pricing is $0.042 per million input tokens, and output tokens are free.
For comparison, The Register reported that OpenAI’s GPT-5.6 Terra costs $2.00 per million input tokens and $12 per million output tokens. In TypeSafe’s side by side demo, Jev finished in 0.114 seconds for $0.000081, while a chat model took 8.566 seconds and $0.013880.
What this really means is that decisions you could never afford to hand to AI become cheap. By my math, assuming 300 to 600 tokens per message including your questions, sorting 1,000 customer messages costs roughly 1 to 3 cents. You can check every email, every comment, every form entry.
The name is a joke with a point. Jevons paradox says that when a resource gets cheaper to use, people use far more of it. TypeSafe is betting the same happens with intelligence.
Who Can Use Jev (And How)
You will probably never type a message to Jev. Its customers are programs, so most people will meet it through tools built on top of it. Here is how that plays out for different groups.
Everyday users
You do not need to install anything. My read is that if the apps you use start sorting, flagging, and routing things faster, a decision model like this is a likely reason. Think of the app that decides whether your message needs a human reply.
Small business owners and freelancers
Sort inquiries into quote request, support, invoice, and spam. Flag the angry ones. Score new leads from one to five. A developer or a no code automation builder can set this up once, and after that the sorting runs on its own at almost no cost.
Customer support and operations teams
This is the use case TypeSafe leans on most. In its own Choice documentation, a ticket goes to the best matching team, a second team gets a copy if it holds a real share of the probability, and the system asks the customer a clarifying question when confidence is low.
Developers and startup founders
Jev works as a smart if statement. Use it wherever hand written rules are too brittle: routing, scoring, extracting, branching. TypeSafe also pitches it as a checker for other AI outputs, scoring or flagging what a bigger model produced. If you are building agents, it pairs well with the ideas in our post on OpenAI Dots vs Meta Muse and the rest of our Agentic AI coverage.
Marketers and creators
Use Choice and Score to sort comments by topic, rate replies by sentiment, and route direct messages to the right person. This is my application of the question types, not a TypeSafe case study, so test it on your own data first.
Students and researchers
Tag hundreds of survey responses or abstracts by theme, and flag the uncertain ones for you to read yourself. TypeSafe lists large scale classification as a target use, and a Choice question accepts up to 255 options. Treat the output as a first pass, never as the final word.
Game builders and hobbyists
TypeSafe showed a bot playing Doom by sending Jev ten decisions per second, which it says costs about $7 an hour. The bot reads game state as structured data, not the screen. At that price, game characters that react in real time stop being a luxury.
Try Jev in 10 Minutes, No Code Needed
The goal is to see how a sorting decision looks when it comes back with probabilities. Use a decision you make over and over, like triaging your inbox.
- Pick one repeated decision. Example: should this email be answered today, this week, or never?
- Open a place to test it. TypeSafe has a Playground in its console, but new signups were paused in late September, so check whether it is open. The OpenRouter Jev Lab triage demo runs Jev against a chat model on 95 sample support messages, and you need to sign in to run it.
- Paste one real message as the input. Remove names, numbers, and anything private. Do not paste customer data into a new service without reading its terms.
- Add a Choice question. Write the question in plain words, then list two to four options with a one line description each. Make the descriptions clearly different from each other.
- Run it and read the probabilities. Look at the winner, the runner up, and the confidence.
- Repeat with ten different messages. Note which ones come back with low confidence. Those are the messages a human should see.
- Set your own rule. For example, act automatically above a confidence you trust, and send everything else to a person. TypeSafe’s sample code uses 0.3 and 0.5 as thresholds. Pick yours after your own tests.
If the results look good, hand the same question to a developer. It becomes one API call.
Honest Limitations
- It cannot write or reason at length. Jev gives up text generation on purpose. TypeSafe says questions work best as quick gut checks, and complex judgments should be split into smaller questions you combine in code.
- It can still be wrong. TypeSafe says Jev cannot hallucinate, but The Register points out that this is not a like for like comparison. Jev’s answers always fit the format. That does not make them correct.
- No explanation. Heise online, as summarized by Trending Topics, notes that Jev gives no written reason for a decision. The same report describes a developer test where Gemini was slightly more accurate on business emails but ten to twenty times more expensive.
- The big numbers come from the company. The 193.6x faster and 444.6x cheaper figures on TypeSafe’s homepage come from its own workflow tests, and the team says they sit on the high end of real world gains. It also admits it cannot prove its pricing is not subsidized.
- Access has moved around. Jev launched behind a waitlist. Flavio Copes reported that TypeSafe opened access on September 20 and paused new signups two days later because of demand. OpenRouter says Jev is available there with no waitlist, so check the current status before you plan around it.
- Big lists need two steps. A Choice question supports up to 255 options. Larger sets need a two stage approach.
Your Move This Week
Jev is not the next ChatGPT, and that is the whole point. It is a cheap, fast decision maker that sits inside software and handles the sorting no one wants to do by hand. Pick one repetitive decision in your work, write it as a single question with a few options, and test it on ten real examples. You will know within the hour whether this kind of tool belongs in your workflow.
Know someone who keeps asking what Jev is? Send them this post and tell them to try the 10 minute test.
Sources
- TypeSafe AI Emerges From Stealth With $40M in Funding With New Model for Composable AI (Business Wire, September 15, 2026)
- Introducing System One Models and Jev (TypeSafe AI blog, Diogo Almeida)
- TypeSafe AI homepage
- Jev documentation: Introduction (TypeSafe AI)
- Jev documentation: Choice (TypeSafe AI)
- TypeSafe AI debuts model for machines that plays Doom (The Register, September 16, 2026)
- Jev on OpenRouter (OpenRouter documentation)
- Support message triage, Jev Lab (OpenRouter)
- How to get access to Jev and an API key (Flavio Copes, updated October 1, 2026)
- Jev: Warum dieses neuartige KI-Modell gerade Entwickler:innen begeistert (Trending Topics, in German)
- Thinking, Fast and Slow by Daniel Kahneman (Penguin Random House)
- Jevons paradox (Wikipedia)
Leave a Reply