In partnership with
Hey folks,
Every AI launch this year has been the same race. A bit smarter. A bit cheaper. A bit faster at writing you an answer. Astra, Sol, Luna, Opus, Fable. All of them still do the same thing. They generate text, word by word. You wait for it.
Then Jev showed up. It skipped the race. It doesn't write anything at all. Give it a decision, with a fixed set of options, and it just picks one. Instantly. For a fraction of a cent.
It's the first genuinely different idea I've seen an AI company ship in a year and it’s blowing up online.
Let's dive in..
It doesn't write. It picks
Every AI model you've used so far writes its answer out, word by word. Even when the real answer is just "yes." Or "billing department." Or "ignore this one." You're paying full sentence-generation prices for a checkbox.
Jev skips the sentence. You give it a situation. You give it a short list of possible answers.
Is this phishing?
Which department does this go to?
How urgent is this?
It hands back the one it picked, plus a confidence score. All the options get weighed at once. Nothing gets typed out one word at a time.
TypeSafe calls it a "System One model." I'd call it a smart filter, not a chatbot. And that's the whole point. Most of what a business needs from AI is filtering. Not writing.
That's also why it's fast and cheap in a way a chatbot just can't be. 70 to 500 milliseconds. $0.042 per million input tokens. Output is free, because there's barely any output to charge for.
TypeSafe's founder, Diogo Almeida, says: "Models have been superhuman at chat for years, so where is all the automation?"
Worth knowing who's saying it. Almeida isn't a newcomer chasing a trend. He co-invented RLHF and InstructGPT, the training techniques that turned GPT into ChatGPT in the first place. They raised $40 million in seed funding. A $200 million valuation. Out of stealth on 15 September (Forbes).
What people built with it
Built by Ben Tossell
So many examples are going viral in the last week. All with a real number attached. Click through and watch any of them run:
500 emails sorted for 3.5 cents. An entire inbox classified in seconds.
A flight booked in about 7 seconds, for under half a cent. Paired Jev with a browser automation tool to make decisions. The browser carried them out.
724 competitor ads read in 40 seconds for 9 cents. Matt Berman broke down 37 brands by hook, format, offer and call to action.
A live BS meter. An open-source tool called jevmeter It fact-checks a debate or a podcast, sentence by sentence. A full hour-long interview cost under half a cent to check.
The most-watched build of the week, 3.87 million views. Instead of asking a chat model to summarise an AI agent's history, Jev scores every step the agent took. It drops the ones that don't matter. Cheaper and faster than a summary. And nobody had to write one.
A chatbot with no chat model inside it. Jev alone decides which tool to call. Web search. Weather. A to-do list. Nothing generates the reply text. "Responses are instant, no hallucinations," in his words.
Separately, one builder ran it against 348 academic papers. Sorted by topic. 18 cents. 3 to 4 seconds.
None of these are things a chatbot technically can't do. They're things a chatbot is a slow, expensive way to do. You're paying to generate a paragraph when all you needed was a checkbox.
"Can't hallucinate" isn't "can't be wrong"
TypeSafe's pitch is a strong one.
Because Jev only ever picks from a list you gave it, it can never make an answer up. Their own numbers back that up. A flat 0% hallucination rate.
Picking confidently from a list doesn't mean picking correctly. One independent tester checked on something real by sorting phishing emails from legitimate ones. Asked one simple yes-or-no question, Jev got it right 63% of the time. Claude got it right 81% of the time. On this test, the cheap new model was the worse one.
Then the tester tried it differently. Instead of one broad question, they asked Jev 5 specific ones. Does the link use a shortener? Does the sender's address really match the company it claims to be from? That kind of thing. Then they combined the 5 answers, using a bit of their own labelled data. Jev's accuracy jumped to 95%. Ahead of Claude. For a fraction of the price. (Beri.net's full test)
Typesafe means it can't make an answer up. It doesn't mean the answer's right. The fix was never a smarter model. It was a better question, and the data to back it.
Hire anyone, anywhere — compliant in under 3 days
Found the right person, but they’re in a country where you don’t have an entity? Setting one up can take months and significant cost.
Remote removes that barrier by becoming the legal employer through our own entities — handling compliant contracts, local benefits, tax setup, and onboarding for you. In fact, an employee is onboarded to Remote every 7 minutes.
Once they’re hired, the same in-house teams that support employment locally also run payroll — so you’re not bouncing between disconnected providers. Less setup, less complexity, and less time between finding the right person and getting them started.
Get on the wailist
Jev was available last week, but locked down via waitlist this week.
TypeSafe is running this through a waitlist, mine included. So today, this is worth knowing about and watching. Not worth rebuilding a workflow around yet. And when you do get in, the win isn't free. It comes from breaking your question into specific signals, and having enough of your own labelled examples to weigh them properly. Not from asking it the same loose way you'd ask a chatbot.
Cheers, Tim
