I watched a demo of 1,700 emails being sorted almost instantly.
No reply. No summary. No helpful paragraph at the end telling you it was happy to assist.
It just made decisions.
That is the point of Jev, a new model from TypeSafe AI. It is designed to sit inside software and return a bounded answer that code can act on. TypeSafe calls this a System One model.[1][2]
Most of the AI tools people see are writers. You ask a question and get text back. That is useful when you need a draft, an explanation or a plan.
It is not always useful when the real question is much smaller.
Is this enquiry a quote request, a support issue or neither?
Does this message need a callback?
Which queue should this go into?
That is the gap Jev is trying to fill.
What Jev actually does
You give Jev an input. This can be text, a JSON record, or another structured piece of information. You also define the possible decisions it is allowed to make.
It then returns a structured result rather than prose.[3][4]
It can choose one option from a list you define. That is useful for a route such as quote request, support request or other. It can score something against an ordered scale, such as low, medium or high urgency. Or it can return the probability of a yes or no answer. For example: “Does this person explicitly ask for a callback?”[2][4]
The important bit is that Jev does not invent the categories. You set the categories. It returns the best match and the probability behind it.
Here is a simplified example. The email and the values below are fictional. This is the shape of the result TypeSafe documents, not a result from my forthcoming test.[4]
{
"model": "jev-latest",
"answers": {
"email_route": {
"type": "choice",
"choice": "quote_request",
"probabilities": {
"quote_request": 0.87,
"support_request": 0.09,
"other": 0.04
},
"confidence": 0.81
}
}
}
Plain English: the model thinks this is a quote request. It is not certain, but quote request has the strongest probability. A business can decide that a result like this moves into a quote queue, while anything below its chosen confidence threshold stays for a person to check.
That is a better fit for ordinary software decisions than asking a chat model to write a paragraph, then trying to extract one usable answer from the paragraph afterwards.
Why people are paying attention
The launch claim is speed and cost. TypeSafe says Jev evaluates declared questions in parallel rather than generating text one token at a time. It reports a 70 to 500 millisecond response time and pricing based on input tokens, with output tokens unmetered.[1][5]
Those are TypeSafe's figures. They are not my test results.
Greg Isenberg's recent video with Ryan Vogel makes the practical case better than the launch page. Ryan ran Jev across 1,700 of his own emails, asking it to assign a category, priority, spam probability and reply likelihood. He says the run cost 18 cents. That is a useful demo, but it is still one creator's demonstration rather than an independent benchmark.[7]
The interesting bit is not the headline number.
It is the workflow.
You put a small decision model at the front of a busy queue. Clear cases go where they need to go. Uncertain cases go to a person, or to a more capable model that can explain its reasoning.
What it can be useful for
This is where the technology becomes practical.
A contractor gets a mixture of quote enquiries, supplier emails, recruitment messages and support requests. Most of them still land in one inbox. Somebody has to read every one before deciding what happens next.
Jev could make the first pass.
It could sort an enquiry into a route. It could flag the messages that need a reply. It could score the likelihood that a lead is worth following up. It could decide whether an incoming job should be reviewed by a person now or added to a standard workflow.
The same idea works for support tickets, document queues, review checks, content libraries and agent workflows. Vercel's AI Gateway lists Jev for choices, scoring and Boolean evaluation inside software.[6]
Think of it as an AI traffic cop. It does not write the reply. It decides where the reply should go next.
What it cannot do
Jev is not a replacement for ChatGPT, Claude or another general-purpose language model.
It cannot write a quote. It cannot explain a design problem. It cannot answer a client's open-ended question. It cannot invent a useful workflow when nobody has defined the choices or the rules around it.
It is also not a reason to hand over decisions blindly.
A probability is useful only if you test it against real labels in your own workflow. A model saying it is 92% confident does not make it right 92% of the time for your inbox, your customers or your wording. That needs measuring.
Ryan Vogel made the same point in a more entertaining way. He tried using Jev as a Bitcoin buy, hold or sell signal. It did not perform well. His conclusion was sensible: keep it for routing-style work, not portfolios.[7]
How to get it
The direct TypeSafe route is still early access. The company says it is bringing developers off its waitlist, and its quick-start guide says API keys are created through the TypeSafe dashboard.[1][3]
That means a console login does not necessarily mean immediate API access.
There is also an alternative route through Vercel AI Gateway, which announced Jev availability on 16 September. That route has its own account and access requirements, so it is not a magic bypass. It is simply another way to use the model when available.[6]
The right way to start is not to connect it to a live inbox and hope for the best.
Start with invented data.
The test I am running next
I am going to test Jev on a small fictional email set first.
Thirty messages. Three routes: quote request, support request or other. A second yes or no question: does the person explicitly ask for a callback?
The labels will be written before Jev sees the messages. Then I will compare its choices, probabilities, confidence, latency and actual cost against those labels.
No customer data. No automatic actions. No hiding the mistakes.
That is the only sensible way to see whether it is useful.
Over the next week, I will get access sorted and run the test. The follow-up post will show the results, including where Jev is uncertain and where it gets the answer wrong.
That is more interesting than another launch-day demo.
Image credits
- Diogo Almeida, founder of TypeSafe AI: Official portrait from TypeSafe's team page at typesafe.ai/team. Used for editorial context under fair use / with attribution. Source: TypeSafe AI Team.
Sources
[1] https://typesafe.ai/blog/introducing-system-one-models-and-jev [2] https://docs.typesafe.ai/introduction [3] https://docs.typesafe.ai/introduction/quickstart [4] https://docs.typesafe.ai/api [5] https://docs.typesafe.ai/models [6] https://vercel.com/changelog/typesafe-ai-jev-now-available-on-ai-gateway [7] https://www.youtube.com/watch?v=4mTLpuQpB80