TypeSafe AI came out of stealth on 15 September with $40 million in funding and a model that does not generate text. You hand Jev a block of state and a list of typed questions. It answers all of them at once and returns each answer as a value with a probability attached.
TypeSafe claims Jev is 193.6x faster and 444.6x cheaper than frontier LLMs, answers in well under a second, and never hallucinates or returns a type error. The first independent benchmark puts the speed and cost gains closer to 3x and 12x. The hallucination claim holds only in a narrow sense, because Jev never writes anything. It spreads probability across answers you defined in advance, and nothing in that stops it putting the weight on the wrong one.
#What it is
Jev is what TypeSafe calls a System One model, after the fast half of the split in Daniel Kahneman's Thinking, Fast and Slow. System 1 is the snap judgement you make without noticing you made it. System 2 is the slow, effortful one. Jev is meant for the small decisions software makes over and over: ranking, scoring, picking between options.
A frontier LLM does that as a side effect of writing text. Jev takes the writing out. You give it the list of answers it is allowed to give, and it tells you how likely each one is.
A request carries state and questions, and questions come in three primitives. A noul returns the probability that a statement about the state holds. A choice picks one of up to 255 named options and returns a probability for each. A score rates the state against ordered levels. Here is an email-triage request adapted from TypeSafe's docs:
{
"model": "jev-latest",
"input": {
"state": "From: IT Helpdesk <no-reply@acme-sec-verify.top> ...",
"questions": {
"action": {
"type": "choice",
"instructions": "What should the mail gateway do?",
"criteria": ["deliver", "quarantine", "reject"]
}
}
}
}{
"answers": {
"action": {
"type": "choice",
"choice": "quarantine",
"confidence": 0.71,
"probabilities": { "deliver": 0.06, "quarantine": 0.71, "reject": 0.23 }
}
}
}In this response confidence is the winning option's probability, so a peaked distribution reads as sure and a flat one as unsure. Input costs $0.042 per million tokens and output is free. The context budget is 64k for state and questions together. TypeSafe describes the model as transformer-based, trained on synthetic data with a method it calls Reinforcement Learning for Calibrated Decisions, but has not yet published a paper, weights or a parameter count.
#The same job, done two ways
When you ask an LLM to triage that email, it starts writing. It spells out {"action": "quarantine"} token by token, your code parses the string back into a value, and if the model added a line of explanation or wrapped the JSON in a code fence, the parse fails and you pay for the call again. Jev gets the three allowed options up front and returns a weight for each, so the answer cannot come back in any shape other than the one you asked for.
The speed comes from not writing token by token. The type safety comes from the answer being picked rather than written, which is how the KDnuggets piece on the launch reads the claim: zero hallucinations means zero out-of-schema outputs. The flip side is that Jev cannot tell you why it quarantined the mail. It has no way to produce an explanation.
#Three things you would build with it
From TypeSafe's documented patterns:
Intent routing. Classify what the user wants and hand it to the right handler. A message comes in, Jev decides whether it is a refund, a bug report or a sales question, and your code passes it on. This pays off only when the routing decision costs less than the work it routes.
Confidence-gated routing. Act on high confidence, ask the user in the middle, hand it to a person or a bigger model when it is low. Your agent is about to delete a file. Jev is 95% sure that is what the user asked for, so the agent goes ahead. At 60% it stops and asks first. The docs warn against one global threshold: reading a file can tolerate doubt that moving money cannot.
Speculative fan-out. Questions are answered in parallel and billed on input only, so ask everything you might need in one call and let your code decide what mattered. One pass over a support message can ask whether it is a refund request, whether the customer is angry, and whether an order number is present. You pay for reading the message once, not three times.
#Do the numbers hold up?
TypeSafe does not score Jev against a standard benchmark. It compares Jev to GPT-6 Astra and Fable 5.1, and treats the average of what those two returned as the right answer. The eval therefore measures how often Jev agrees with two frontier models, which is a different number from how often it is right. TypeSafe says as much, and adds that its own team wrote the test workflows and that the headline gains are the high end of what it measured.
The better evidence so far is jev-phishing-bench, a reproducible run over 2,000 emails published two days after launch. Asked outright whether an email was phishing, Jev scored 62.6% and Claude Haiku 4.5 scored 81.3%. Split into five narrow signal questions, with a logistic regression fitted on 1,000 labelled emails, Jev reached 95.0% and Haiku 93.2%, a gap the run's own significance test could not separate from chance (p = 0.063).
Jev is weak on the one big question and strong on the small ones, which matches what TypeSafe's docs recommend: break a hard question into several easy ones and combine the answers in your own code. That combining step is a model of its own, though, and you have to fit it on your own labelled data before any of it works.
Speed and price from the same run: 239ms against Haiku's 687ms, and $0.038 per thousand emails against $0.462. That is about 3x faster and 12x cheaper. A real gain, and a long way from 193x and 445x.
Two results from the same run cut the other way. On expected calibration error Jev scored 0.154 and Haiku 0.097, so on the only public test of the property TypeSafe leads with, Jev was the less calibrated of the two. A plain regex over the links and the sender scored 91.6%, far above Jev's 62.6% when asked the phishing question outright. Aman Kumar's 16,000-call evaluation shows where it tops out as the options multiply: accuracy slides from 98.7% on two-option spam to 76.0% across Banking77's seventy-seven intents.
That benchmark has soft spots. The emails are synthetic, and the phishing labels come from a check on the links rather than from anyone reading the mail. Most of the signal either model could use sat in the links and the sending domain. These are the first public tests of a model that has been out for a week. Better ones will follow.
#So, new toy or breakthrough?
Classification models are not new. We have been sorting text into buckets with small, fast, cheap models for years, and the LLM frenzy made a lot of teams forget that anything else exists. What Jev changes is the price and the reach: an order of magnitude cheaper than asking an LLM, over questions you write in plain language instead of a model you train per task. It is also very new. Users got access a few days ago, the API is closed, there is one outside benchmark, and nobody without a stake has confirmed the calibration claim. Enough to put Jev in front of an expensive model. Not yet enough to put it in place of one.
AKENA is an engineering studio working across AI, blockchain, and software infrastructure. We take projects from first principles to production and operate what we ship. If you are weighing a new model against the one you already run and want the comparison measured on your own data rather than on a vendor's, we should talk.