TL;DR
Jev is a new AI model from Typesafe AI, and it does not write text. It looks at a piece of information and a set of possible answers, then returns a fast, structured decision with a confidence score attached. Typesafe calls it a "System 1 model": built for split-second classification and routing, not conversation. That narrow focus is exactly why it is spreading so fast right now. Most small businesses already have a pile of decisions shaped exactly like this: which lead is real, which ticket is urgent, which call actually matters. Until now, making that call meant paying full chat-model prices just to sort things.
What Jev actually is
Typesafe AI announced Jev on September 15, 2026, as the first of what it calls "System 1 models," a name borrowed from psychologist Daniel Kahneman's idea of fast, intuitive thinking, as opposed to the slow, deliberate kind chat models are built for. Instead of generating a response one word at a time, Jev takes a block of information and a set of typed questions, then answers all of them in a single pass. No back-and-forth, no drafting.
It is trained with a method Typesafe calls Reinforcement Learning for Calibrated Decisions, and every answer comes back as a typed value with a confidence score attached, not a paragraph of prose. Typesafe's own framing is that this makes hallucination close to impossible for the outputs it actually returns, since the model cannot claim a category that does not exist. The categories are fixed ahead of time. It can be uncertain. It cannot make things up.
The published numbers are the other reason people are paying attention. Typesafe claims 70 to 500 millisecond response times, roughly 40 to 200 times faster than a comparable large language model on classification-shaped tasks, at $0.042 per million input tokens with output tokens free. In one of the company's own examples, sorting 1,700 emails into junk, needs-a-reply, and can-wait cost 18 cents, total.
Why it is suddenly everywhere
Three things happened close together. Typesafe shipped a free, fully open-source Chrome extension built on Jev that replaces browser find-on-page with semantic search: type what you mean instead of the exact text, and it still finds it, in something close to real time. Developers started swapping it into existing products almost immediately. According to TechCrunch's coverage, Vercel replaced an existing LLM call with Jev in one workflow and saw responses come back five to eighteen times faster, with better accuracy on top.
Then a handful of builders started publishing exactly what they had swapped out. One widely shared breakdown showed seven SEO workflows (page scoring, citation checks, internal link mapping) that had quietly been running on a full chat model and did not need to be. Another showed the "1,700 emails for 18 cents" example above, framed around a simple idea: most inboxes, ticket queues, and application piles are not asking an AI to write anything. They are asking it to decide, over and over, as fast as possible.
5 ways small businesses can use a classifier model like this
Jev itself is in limited early access, gated behind a waitlist. But the pattern behind it, a small, fast model that only has to decide and not write, already fits a long list of things most small businesses currently do by hand, or overpay a chat model to do. Five of the clearest:
- Instant lead scoring. Every contact form and quote request gets read and scored the moment it arrives: real inquiry or spam, urgent or not, before a human ever opens it.
- Support ticket triage. A ticket gets classified and routed to the right queue before anyone reads it, which matters most for teams too small to have a dedicated support desk.
- Call and message classification. The same scoring idea applies to phone calls: is this caller a returning customer, a new lead, an emergency, or a wrong number, decided in real time instead of after the fact.
- Application and intake screening. Job applications, permits, insurance claims, and loan documents all get read and sorted one at a time today by someone on staff. A classifier does the sorting; a person still makes the actual decision.
- Inbox and content triage at scale. Whether it is 1,700 emails or a queue of reviews, the same "which of these actually need a human" sort applies, at a cost low enough to run continuously instead of once a day.
Every business's version of this looks a little different. Pick what is closest to yours:
Interactive
What would this look like for your business?
Pick what's closest to what you do, then select the ideas worth building and we'll follow up.
What this would actually cost you
The numbers above are Typesafe's own, so they are worth checking against something concrete: your own volume. The calculator below uses Typesafe's published pricing and its own stated multiple over a typical LLM call, next to a rough estimate of doing the same sorting by hand.
Interactive
What would this cost at your volume?
Enter how many things you'd classify a month (leads, tickets, emails, whatever it is) and see roughly what a Jev-style classifier costs next to a typical LLM call and a person doing it by hand.
Jev-style classifier
$0.53
~21 min of processing, total
Typical LLM call
$106
$105 more than Jev
Doing it by hand
$833
$833 more than Jev
Rough estimate, not a quote. Jev cost is based on Typesafe's own published example (1,700 emails classified for 18 cents) and its stated "up to ~200x cheaper than a comparable LLM" claim for classification-shaped tasks. Human cost assumes a $20/hr wage. Actual numbers depend on your data and workflow.
Where this connects to your phone
An inbound call is the same shape of problem, just spoken instead of typed. Is this caller a hot lead, a returning customer, a spam call, or an emergency that needs to jump the line? That gets decided while the phone is still ringing, not after someone finally calls back. It is the same classify-then-act pattern Jev is built for, and it is also, more specifically, what an AI receptionist does for a business's phone line every time it rings.
Common Questions
Is Jev open source?
The Jev model itself is proprietary and only accessible through Typesafe's API. The Chrome extension built on top of it, a semantic find-on-page tool, is open source.
Is Jev available to everyone right now?
No. Jev is in early access, and Typesafe is bringing developers on from a waitlist gradually rather than opening it up all at once.
Is Jev the same kind of AI as ChatGPT or Claude?
No. ChatGPT and Claude are built to generate open-ended text and hold a conversation. Jev does not generate text at all, it returns typed answers to predefined questions. Typesafe designed it to sit alongside a chat model in a workflow, not replace one for writing tasks.
Do I actually need Jev, or would a regular LLM prompt work fine?
If the task is genuinely open-ended, like writing an email or answering a novel question, a general-purpose LLM is still the right tool. A classifier model like Jev is worth it once the task has a fixed, known set of possible answers and you are running it often enough that speed and cost start to matter.