Jev vs. ChatGPT: the AI that decides instead of talks
A new AI model called Jev doesn't write a single word. It makes fast, cheap decisions with a confidence score. Here's what that means for a trades shop.
Meet the AI that refuses to talk
On September 15, a company called TypeSafe AI released a model named Jev, and a corner of the internet lost its mind. So many developers piled in that TypeSafe's servers briefly couldn't keep up. The founder, Diogo Almeida, spent four years at OpenAI working on ChatGPT. Now he's built something that does almost none of what ChatGPT does.
Jev can't write you an email. It can't answer a question in a sentence. Ask it to explain itself and you get nothing. It doesn't produce words at all.
It makes decisions. Fast, cheap, and with a number attached that says how sure it is.
That sounds like a downgrade. For a lot of the work inside a trades shop, it might be the opposite.
First, what ChatGPT actually does
ChatGPT, Claude, and Gemini are all large language models, or LLMs. Under the hood, an LLM does one thing over and over: it predicts the next word. Then the next one. Then the next one. A paragraph is a few hundred of those guesses stacked up, one at a time.
That's why ChatGPT is so good at writing the apology email to the customer whose tech showed up three hours late. Writing is guessing the next word, done really well.
It's also why ChatGPT is slow and expensive for small, boring decisions. If you ask it "Is this call a booking or a complaint?", it still has to generate an answer word by word. It might say "Booking." It might say "This appears to be a booking, although the customer does mention frustration with a previous visit." It might, on a bad day, invent a third option you never offered. Then somebody has to write code to read that answer and figure out what it meant.
What Jev does instead
Jev doesn't predict words. You hand it two things: the situation (a call transcript, a review, a tech's job note) and a question with a fixed set of answers. It hands back a probability for each answer.
Ask it "Booking, complaint, emergency, or sales pitch?" about a call where somebody says water is coming through their kitchen ceiling, and you'd get back something shaped like this:
| Answer | Jev's confidence |
|---|---|
| Emergency | 94% |
| Booking | 5% |
| Complaint | 1% |
| Sales pitch | 0% |
That's the whole output. No paragraph. No "Great question." Just the call, sorted, with a confidence score.
This is where the company name comes from. In software, "type-safe" means the answer is always the kind of thing you asked for. Give Jev four options and it answers with one of those four options, every time. It can't say "banana." It can still be wrong (it's a prediction, and TypeSafe says so), but it can't go off-menu.
Your best dispatcher already works like this
TypeSafe calls Jev a "System One" model. That comes from the psychologist Daniel Kahneman, who split thinking into two speeds. System One is the fast, gut-level call. System Two is the slow, careful, write-it-out kind.
You've watched both happen at your dispatch desk. A caller says "I smell gas" and your senior dispatcher is moving before the sentence is over. Nobody writes a memo about it. That's System One. Then the same dispatcher spends twenty minutes drafting the message to a customer whose install got pushed a week. That's System Two.
ChatGPT is built for System Two work. Jev is built for System One. Most of the AI hype has been about the memo writer. Jev is the first big release aimed at the gut call.
The numbers that got everyone excited
Fair warning: some of these come from TypeSafe itself, and they say plainly that their own team built the tests.
- Jev answers in 70 to 500 milliseconds. DigitalOcean's write-up puts the big LLMs at 3 to 329 seconds for the same kinds of jobs.
- On TypeSafe's own workflow tests, it ran 40 to 200 times faster and 40 to 400 times cheaper than frontier LLMs.
- Its posted price is $0.042 per million tokens of input, and the output is free. (A token is roughly three-quarters of a word.)
- An outside benchmark by DataCamp had Jev and GPT-5.6 Terra basically tied on sorting text by sentiment, 67.8% accurate versus 67.9%, with Jev costing about 1/76th as much per decision.
Put that price in shop terms. Say a call transcript runs about 1,500 words, call it 2,000 tokens. A shop taking 30,000 calls a year is feeding in 60 million tokens. At $0.042 per million, sorting every single call you take all year costs about $2.52.
Two dollars and fifty-two cents. For the year. That's the part that made developers sit up.
Why it's named after a coal economist
This is our favorite fact about Jev, so bear with us.
The model is named after William Stanley Jevons, a 19th-century British economist. In the 1860s, steam engines were getting much more efficient with coal, and everyone assumed England would burn less of it. Jevons pointed out the opposite was happening. Cheaper power meant people found a hundred new things to power, and coal use went up.
That's called the Jevons paradox, and TypeSafe named their model after it on purpose. Their bet is that when a decision costs a fraction of a cent, people stop asking "is this worth automating?" and start automating everything.
For a trades shop, that's a real shift. Today, nobody runs AI over every single call, review, and job note, because it costs too much and takes too long. At $2.52 a year, the math stops being the reason not to.
Where this shows up in a trades shop
Think about the small sorting decisions your office makes hundreds of times a day:
- Is this call a new booking, a complaint, an emergency, or a vendor selling something?
- Does this booking have the right job type, or will dispatch kick it back on Thursday?
- Is this five-star review actually a two-star review with a typo? Is the one-star review about a tech, a price, or a scheduling miss?
- Does this tech's job note mention something unsafe that a supervisor needs to see today?
- Is this customer likely to cancel their membership this month?
None of those need a paragraph. They need a fast, honest answer and a way to know when the answer is shaky. That's the same kind of decision our Call Scout makes after every call and Data Cleanse makes on every booking. Models like Jev are about to make that kind of sorting dramatically cheaper for every software company in the industry.
What Jev still can't do, by design, is write the text that tells the customer their tech is running late, explain to the owner why the margin dropped, or hold a conversation. For that, you still want ChatGPT, Claude, or a person.
The part that matters most: the confidence score
The most useful part of Jev isn't the speed. It's that every answer comes with an honest number for how sure it is.
That number is what lets you set a rule like this one: if the AI is 90% sure a call is an emergency, page the on-call tech. If it's 60% sure, put it in front of a CSR. A tool that can tell you when it isn't sure is a tool you can actually trust with your phones.
A chatbot will write you a confident paragraph whether it's right or wrong. That's a big reason owners get burned by AI.
So what should you do?
Nothing this week. You won't be signing up for Jev yourself. It's a building block for software companies, and it's in limited early access.
What you'll see is the software you already use getting faster and cheaper at sorting things. When a vendor pitches you on "AI that triages your calls" or "AI that flags risky jobs," ask them two questions:
- When your AI sorts something, how sure is it, and can I see that number?
- What happens when it isn't sure?
If they can answer both, they built it right. If the answer is "it's AI, it just knows," keep your wallet in your pocket.