Routing 10,000 support tickets a month through the Jev AI model would cost about $0.21. That’s our arithmetic, using TypeSafe’s early-access price and a guess of 500 tokens a ticket. The price isn’t the problem. The problem is the word “hallucination”, which TypeSafe uses to mean the answer always fits the format, and a tidy wrong answer is harder to catch than a messy one. Test its confidence on 200 of your own cases before you wire it into anything.
TypeSafe, a San Francisco startup whose CEO is a former OpenAI researcher, exited stealth on September 15 with a $40 million seed round led by DCVC, per VKTR. Jev is in early access, so you’re joining a waitlist, not buying a product.
Key points
- What does the Jev AI model actually do? Jev doesn’t write anything. You send it a state, such as a support ticket, plus typed questions, and it returns a probability or a choice with a confidence value for each, all in one parallel pass.
- Does zero hallucination mean the answers are right? No. It means the output always matches the schema, so Jev can’t return a value outside your list or a malformed answer.
- How do you test Jev’s confidence in an afternoon? Pull 200 past tickets where you already know the right routing.
- What does it cost to route 10,000 tickets? Almost nothing. Say a 40-person firm sends 10,000 tickets a month at roughly 500 input tokens each.
What does the Jev AI model actually do?
Jev doesn’t write anything. You send it a state, such as a support ticket, plus typed questions, and it returns a probability or a choice with a confidence value for each, all in one parallel pass. Think of it as a very fast classifier you configure with plain questions, not a chatbot.
TypeSafe’s launch post describes a model built to make fast, structured decisions that software can use directly. The question types, as laid out in eesel AI’s explainer, are yes or no statements, picking one option from a list, and scoring against a rubric. The context window is reportedly 32,000 tokens. It won’t generate text, explain itself, reason in steps or hold a conversation.
It decides. It doesn’t act. Refunds, ticket updates and escalations stay in your own pipeline, which is where you’d want them anyway.
Does zero hallucination mean the answers are right?
No. It means the output always matches the schema, so Jev can’t return a value outside your list or a malformed answer. That’s a real property and a useful one. It isn’t accuracy. TypeSafe’s own post says the 0% figure “is not empirical” because conformance is guaranteed by design, and the company hasn’t published calibration metrics.
Here’s why a buyer should care. You route support tickets to billing, sales or legal, and Jev says billing at 0.95 when the right answer was legal. The system has made a confident, well-formed, wrong decision, and nothing about it looks off. We’d rather catch ten garbled answers than one clean wrong one. A garbled answer announces itself. A clean wrong one sits in the billing queue until a customer complains.
The word to learn here is calibration. A calibrated model’s 90% confidence is right about nine times in ten. An uncalibrated one can say 90% and be right half the time, and it says it just as smoothly. Nobody outside TypeSafe has measured which one Jev is. That’s the whole gap between the marketing and the buying decision.
We’ve seen the same kind of vendor claim before, in our piece on GPT-5.5 Instant’s hallucination numbers. For what a wrong decision costs you afterward, see our AI agent incident plan, and for the permissions side of automated decisions, our AI agent security checklist.
How do you test Jev’s confidence in an afternoon?
Pull 200 past tickets where you already know the right routing. Send them through Jev and record the answer and the confidence for each. Then sort by confidence and check whether the 0.9 group is right about 90% of the time. Big gaps mean you shouldn’t automate on its scores.
Two details make the test honest. Include the ugly tickets, the ones that mention billing and legal in the same message, because that’s where a confident wrong answer lives. And pick the threshold before you look at results. If you decide afterward that anything above 0.8 is “good enough”, you’ve graded yourself on a curve.
If Jev passes, route automatically above a confidence line and send the rest to a person. If it fails, you’ve spent an afternoon, not a quarter. Either result is cheap next to finding out from customers.
Keep the 200 cases afterward. When TypeSafe ships a new version, or the early-access price changes, rerun the same set and compare. A test you can repeat in ten minutes turns a vendor’s upgrade announcement from a leap of faith into a diff. Put one person’s name on it, and put the date of the last run in the same spreadsheet. Before you add any new model, run it through an AI vendor safety risk checklist.
What does it cost to route 10,000 tickets?
Almost nothing. Say a 40-person firm sends 10,000 tickets a month at roughly 500 input tokens each. That’s 5 million tokens. At TypeSafe’s list price of $0.042 per million input tokens, with output free, the monthly bill is about $0.21. The token counts are our illustration, and the price is TypeSafe’s early-access rate.
Responses, per the vendor page, take 70 to 500 milliseconds. Vendor benchmark tables need the same scepticism, so test Kimi K3 and Jev on your own tasks.
TypeSafe claims Jev is 193.6x faster and 444.6x cheaper than frontier models on its own workflows, and says those are probably high-end results. Its homepage puts the input price 238x lower than Claude Fable 5.1. Treat every one of those ratios as marketing. TypeSafe also admits it can’t prove its pricing isn’t subsidized, so a price that low may not last.
A sticker price is rarely the whole bill, which is why we argue for measuring cost per successful task. A wrong route that costs a customer is worth more than every token Jev will ever bill you. Our guide to AI costs for Canadian business covers the wider budget.
How much of this is TypeSafe grading its own homework?
Most of it. TypeSafe says its capabilities team built the workflows, the reference answers were the average of GPT-6 Astra and Fable 5.1 outputs, and speed tests ran from laptops on the West Coast. Each of those can tilt results in Jev’s favour, and to its credit the company says so.
A claim from Vercel’s CEO that Jev was up to 18x faster at the 95th percentile for a safety reviewer comes via TypeSafe and hasn’t been independently confirmed. There’s also a cap of 255 options per question, with occasional slowdowns above that.
The sceptic’s line is that Jev is a classifier with good marketing, and that’s partly right. Small, fast models that sort things into buckets have existed for years, and a well-tuned one costs next to nothing too. What’s new is the interface. You get a confidence-scored answer by asking a plain-language question without training anything.
So the call depends on your shop. If you already have a classifier that works, skip Jev. If you have a pile of messy routing rules and nobody to build a model, give it the 200-ticket afternoon. We’d change our caution if someone published reliability curves on public datasets showing the confidence scores track accuracy. Until then we’re guessing about as much as TypeSafe’s marketing is.
Frequently asked questions
Can the Jev AI model really not hallucinate?
Only in the sense that its answers always match your schema. It can still pick the wrong option and report high confidence, and TypeSafe’s post says the zero figure is not empirical.
What would a small business use Jev for?
Narrow decisions that code consumes, such as routing tickets, classifying messages, flagging policy issues or gating other steps. It doesn’t write text or take actions.
How do I test whether Jev’s confidence can be trusted?
Run about 200 past cases where you already know the right answer. Group Jev’s results by confidence and check whether the 0.9 group is right close to 90% of the time. Large gaps mean don’t automate on its scores.
Archive entry dated September 16, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
