Jev doesn't write. That's the feature.
You have a Jev API key and a blank file. The useful answer is what comes back from a call: a typed choice, a score, or a yes/no probability. It does not hand you a paragraph.
I have not run Jev on my own data. I read the TypeSafe docs, the TechCrunch piece, and the Hacker News thread, because the landing pages do not agree with each other. Treat this as a map, not a build log. If you want the kind of post that comes after a real integration, that is a later piece. This one gets you oriented, with a practical eye on where a tool like this would sit in a real system.
What a call returns
When the next line of your program is a branch, chat models can feel like the wrong tool. You often want a closed decision with a confidence score, not a long stream of prose. Jev is TypeSafe's System One model aimed at that gap. You hand it a state (plain text, JSON, or an array) and questions whose answers you already listed. It returns a typed decision with a probability. Your code decides whether to act, and where the confidence threshold sits.
There are three question types. Choice picks one option from your list and shows the probability of each. Score places the input on levels you ordered, like calm up through ready to leave. Noul is a yes or no: a probability from 0 to 1 that one specific statement is true. Some writeups call that a boolean. Same idea.
You can ask several of those about one state in a single call. They are meant to stay independent. "Which team owns this ticket?" does not answer "how angry is the customer?" Do not let one stand in for the other.
The System One name is Kahneman's split: fast intuition versus slow reasoning. TypeSafe is selling the fast side. For a developer the useful bit is smaller than the psychology talk. The question needs a closed set of answers, and two reviewers should be able to agree on what a good one looks like. "Handle this incident" fails that test. "Which of these teams should look first?" passes.
Diogo Almeida, who worked on ChatGPT and the RLHF paper at OpenAI, left and started TypeSafe on this bet: chat is the wrong shape for software. A program wants a branch. The weights are not public. The training recipe, which they call reinforcement learning for calibrated decisions, is still mostly a name. When the homepage says the model cannot hallucinate, read that as "it cannot emit a label you did not allow." It can still be wrong on your ticket. The shape of the answer is constrained. The truth of the answer is still your problem.
A legal label can still pick the wrong team
"No hallucinations" means Jev will not invent a team name you left off the list. It can still send the payments ticket to storefront. If the report never says whether the request reached the payment service, a forced choice between payments and storefront hides the hole. When the workflow needs an out, put "not enough evidence" in the options.
Confidence is not the same as "I am sure." In TypeSafe's own docs example, billing can win at 0.84 while confidence sits near 0.6, because another option still holds a real slice of the probability. Put the threshold in your code, next to what a wrong action costs: auto-act only when confidence is high, otherwise review or hand it to a person. Finding the ticket owner is not permission to let Jev roll back a deploy.
I care about that line because I have shipped production RAG and agent systems. A wrong branch still lands on a real user. Jev does not remove that cost. It only makes the branch cheap to ask for.
The numbers, with the fine print
Figures below are as reported in the sources I read, up to 2026-10-03. None of these are my benchmarks. If you already meter credits or rate limits in a gateway, the useful question is how the numbers look under your own traffic.
The giant headlines, on the order of 200 times faster and 400 times cheaper, come from TypeSafe's own workflow evals on classification-shaped jobs, not writing. They are easy to read as a promise that you will see the same gains. MarkTechPost keeps the footnotes the homepage skips. TypeSafe wrote the workflows. The reference answer is an average of two frontier models. The company says those gains sit at the high end of real use, and that it cannot prove the price is unsubsidized.
Outside that spreadsheet the numbers shrink, and they are more useful for planning. A Vercel engineer swapping Luna for Jev on a command-safety classifier saw roughly 5 to 18 times faster. An email-sorting test found Gemini a little more accurate and 10 to 20 times more expensive. On a bounded question it can be faster and cheaper. You still measure quality yourself.
The price in that MarkTechPost breakdown is $42 per billion input tokens, which is $0.042 per million. Output tokens are free. A few community pages say $0.42 per million instead, so use the number on your invoice. Latency people quote from the docs is roughly 70 to 500 milliseconds.
Where to put the key
The API shape is small: state in, questions in, typed answers out. Start with the TypeSafe docs. Guides also show Vercel AI Gateway, and forum threads mention OpenRouter. Stick to TypeSafe's docs or a gateway you already pay for. Lookalike playgrounds say in the footer that they are someone else's product.
Repos from the first week with a concrete job (list as of 2026-10-03). Timings are the authors' numbers.
If you just want it inside the agent you already run:
- typesafe-mcp puts Choice, Score, and Noul on Claude Code, Claude Desktop, Codex, and Pi.
- jev-mcp is the judgment kit: fact-check, screen content, rank, classify, extract.
Where speed actually matters is lots of tiny decisions inside a loop:
- jev-ultrafast is a browser agent. Each step, Jev only says what to click. A small language model is called when something has to be typed. Their Google Flights clip is about 7 seconds and roughly $0.004.
- fast-jev-compaction and Winnow both sit on Claude Code and decide which tool output is still about the task. What they keep stays verbatim. They do not rewrite it into a summary.
- Blink walks a repo one folder at a time and asks which files are closer to the current issue.
Narrower tools for when your case is already clear:
- json-render, from Vercel labs: the idea going around is that Jev picks the component, the props, and the layout, instead of writing JSON token by token.
- SemDecide is the shell version, for a crawler, CI, or a data pipe.
- jev-codex-router asks how hard this coding turn is, then picks the model tier.
- jev-review flags the risky parts of a diff before a more expensive model, or a person, reads the rest.
shipwithjev.com is a community catalog if ten links is not enough. TypeSafe's own site is typesafe.ai.
What it does not do
Jev does not write the incident update, the email, or the patch. It also does not look at an image or an audio clip until you turn that into text. The docs treat English as the language it is best at. For anything else, test it and watch the confidence.
Use it when the next line of your program is a branch you are tired of hand-writing, and the input is already a mess of language: ticket owner, tool-call risk, whether this tool result is still relevant, whether this turn is done. If you need a paragraph, call the model that writes paragraphs, and let Jev decide whether that call is worth making.
Related: WTF is RAG anyway?