Ryan Vogel just classified 1,700 emails for 18 cents.
Category, priority, spam score, and whether each one deserves a reply. The run chewed through 4.2 million input tokens and 500,000 output tokens. Total bill: $0.18.
The model behind it is called Jev AI, and it is not an LLM. You can't chat with it and it never streams a wall of reasoning at you. It takes your data plus a list of possible answers and returns a probability for each one in about 200 milliseconds.
Greg Isenberg brought Ryan (founding team at OpenCode) on his pod the day Jev dropped to break down what it is, where it wins, and which startup ideas it just unlocked. I pulled the numbers out of the whole episode so you don't have to.
What Jev AI Actually Is (in One Example)
Forget everything ChatGPT trained you to expect.
Jev is a classifier. You define an input and an output schema, and it returns a probability for every option in that schema.
Ryan's example: hand it a photo description of an iPhone and ask "what color is this?" with blue, orange, red, green, and yellow as the options. Jev doesn't write a paragraph about the phone. It comes back with something like 80% orange, 10% red, 10% blue.
That's it. That's the whole model.
→ Input goes in, a decision with confidence scores comes out, and you never wait for tokens to stream.
Ryan calls it a decision model. Greg's mental model is even better: an AI traffic cop. Information arrives, Jev decides what it is, how important it is, and where it goes next.
The 18-Cent Email Demo
The demo that sold me: Ryan ran Jev over 1,700 of his real emails.
Each email object (subject, body, sender, the works) went in as raw input. Four outputs came back per email:
- Category: shopping, work, marketing, finance, security
- Priority: five levels, from low to urgent
- Spam score: a 0 to 1 scale, not a yes/no
- Reply percentage: how much this email deserves a response
One user email flagged at 90% on the reply scale. It was a customer reporting a possible account violation. Exactly the kind of message that dies in a founder's inbox for three days.
Ryan set up the run expecting it to take hours. It finished while he was still complaining about how long it would take, on camera. 4.2 million input tokens, 500,000 output tokens, 18 cents.
For comparison: piping 1,700 emails through a frontier LLM to score them one by one would cost real money and take real time. Jev answers each query in roughly 200 milliseconds regardless of the schema.
It Doesn't Generate Text. That's the Point.
Here's the part that rewired my brain a little.
Jev's spec technically generates no text at all. The schema you pass in defines every possible output. The model just assigns probabilities to your options and hands back a clean, type-safe object you can drop straight into code.
You never parse a chatty response or strip a "Sure! Here's your JSON:" wrapper, and it can't hallucinate a category you didn't define.
Ryan even stress-tested the boundary. He built a toy where Jev "types" letter by letter, choosing each next character A through Z as a separate decision. Asked "what is bigger, a cat or an elephant?", it slowly spelled out an answer. Janky, but it made the point: Jev is a decision engine, and the chat interface simply isn't there.
If you've been following how founders wire agents together (I broke down Ras Mic's system in the software factory piece), Jev slots in as the cheap, instant decision layer those pipelines keep faking with slow LLM calls.
Cheap Enough That Price Stops Being a Question
The OpenCode team got a $5 intro credit when they set up their account.
They hammered it for two days straight, every demo in this episode included, and didn't burn through it. Ryan's estimate: load $10 and you're probably set for three months of experiments.
The first real business use case in the episode is already running. Ryan's girlfriend runs a graphic design agency with a contact form, and like every agency, the inbound is a mix of dream clients and solicitation spam.
Jev now scores every submission on a 0 to 1 "is this a good lead" scale. A 98% lead gets a fast, personal reply. A lukewarm "I think I might want design, not sure" gets a different treatment. That's the entire logic of lead nurturing compressed into a probability score at the top of the funnel.
She didn't build an AI product. She put a 200-millisecond decision in front of an existing queue.
The Startup Framing: Find an Expensive Queue
Greg's whole angle in the episode is one sentence, and it's worth stealing:
→ Find a business with an expensive queue of incoming information, and put Jev at the front of it.
The examples they riff through:
- Support triage. "I need help with X" gets classified and routed to the right team in 200ms instead of sitting in a shared inbox.
- Local services matching. You type "I need my driveway power washed" and Jev scores every provider nearby for fit. Instant match instead of "we'll email you by end of day."
- Actually-instant quotes. Every "get an instant quote" form that isn't instant is a Jev use case wearing a lie.
- Lead mining your own history. Run it over years of old emails and find the high-value clients you never replied to.
None of these need model brilliance. They need speed, price, and a confidence score. High confidence goes to a human. Medium goes to automation. Low gets ignored.
The moat isn't the model (everyone gets the same API). The moat is knowing which queue is expensive in which industry. That's founder knowledge, not AI knowledge.
Where Jev AI Falls on Its Face
Ryan is honest about the limits, which made me trust the rest more.
He hooked Jev up to a Bitcoin signal that decides buy, hold, or sell every minute. His words: "it does not seem to be doing well."
He ran the same test with GPT-6 Astra, OpenAI's frontier model, and it did better because it could cross-reference news. Different tool for a different job (I covered how founders are making money with GPT-6 Astra if that's your lane).
The rule that falls out of it: Jev makes fast, cheap, repeatable decisions on incoming data, in an advisory role. It should not run your stock portfolio, and anything high-stakes still wants a human or a heavier model behind the confidence score.
Two Demos That Hint at What's Coming
Two more moments from the episode that are easy to skip past:
Auto-clipping long videos. Ryan dropped a video file into a 10-minute prototype. It transcribed the file, pushed the word-level transcript through Jev, and scored 17 clip-worthy moments in about 3 seconds across 1.1 million tokens. Every creator paying an editor to "find the good parts" should sit with that one.
Browser control. The browser-use team had Jev drive a browser to book a flight from Zurich to London. Dates selected, flight found, 7.1 seconds. The same task takes a typical browser agent one to three minutes, because every click waits on an LLM to think.
A classifier picking the next click beats a language model writing an essay about the next click, at least for anything you do a thousand times a day.
How to Try Jev Today
Jev is invite-only with a waitlist at typesafe.ai as of the episode's release.
The workaround Ryan shares: Vercel's AI Gateway already has Jev available, so you can start testing immediately without an invite.
His starter move costs you one evening. Describe your business to whatever AI agent you already use and ask: "which decisions do I make daily on incoming data that a classifier could score for me?" Then wire the top answer to Jev with $5 of credit.
Warning from Ryan, who is not selling anything: "It is dangerously addictive."
FAQ
What is Jev AI?
Jev AI is a classifier model from Typesafe AI. You give it an input and a schema of possible outputs, and it returns a probability for each option in about 200 milliseconds. It makes decisions on data (categorize, score, route) rather than generating text like ChatGPT or Claude.
How much does Jev AI cost?
Very little. In the episode, classifying 1,700 emails (4.2 million input tokens, 500,000 output tokens) cost 18 cents total. A $5 intro credit survived two days of heavy demo use by the OpenCode team, and Ryan estimates $10 could cover about three months of experimenting.
How is Jev different from an LLM like ChatGPT?
An LLM generates text and reasons out loud, which makes it slow and comparatively expensive. Jev generates no free-form text. It only assigns probabilities to the output options you define, which is why it answers in around 200 milliseconds and costs a fraction of a cent per query.
What should you not use Jev for?
High-intelligence, high-stakes calls. Ryan's Bitcoin buy/hold/sell experiment performed poorly, while a frontier model did better by pulling in news context. Use Jev for routing, scoring, and triage in an advisory role, not for decisions where being wrong is expensive.
How do you get access to Jev AI?
Direct access is invite-only through the typesafe.ai waitlist. For instant access, Jev is live on Vercel's AI Gateway, and their AI package already supports it, so you can start testing with a few dollars of credit today.
The Founders Who Win Here Won't Be AI People
Every platform shift has a window where knowing a boring industry beats knowing the tech.
Jev is cheap enough and fast enough that the only real question is which expensive queue you understand better than everyone else. Dispatch? Intake forms? Claims? Refund requests?
That's an operator question. Which is exactly why I'd rather you hear how real bootstrapped founders answer it than guess alone.
Every week on the Profitable Founder Podcast I sit down with founders doing $100K to $10M a year and pull apart how they actually found their queue.
