TypeSafe has put a model called Jev behind a waitlist, and its selling point is that the model does not write. Built by a former OpenAI researcher, Jev takes a block of text plus a list of permitted answers and returns one of them with a probability — a label such as "payment problem", and how likely it is that the customer is demanding money back. The company claims responses in 70 to 500 milliseconds. Input costs $0.042 per million tokens and output costs nothing, because in the ordinary sense there is no output. TypeSafe markets Jev as a model that cannot hallucinate.
The documentation's example is a retailer. A customer writes that they were charged twice for an order and asks for the second payment back. Before anyone touches the case it has to be sorted: is this about payment, delivery or returns, and does the customer want a refund or only an explanation? A developer wires the store's software to Jev, supplies those questions and enumerates the acceptable answers. Every new message goes to the model as text, and what comes back is not a reply to the customer but a classification.
The store's own code then applies rules written in advance. Payment questions go to accounting. Refund requests are flagged for further handling. Anything the classification leaves unclear goes to a person. Jev scores; the software decides.
Note what the model does not do. It does not confirm that the double charge happened — that has to be checked separately against real transactions — and it does not return the money. Once the status is established, a separate chatbot can tell the customer. For the triage step itself, no chatbot is needed at all.
The latency claim is where the product argument lives. TypeSafe says Jev is many times faster than even the quickest current language models, because it does not generate text step by step but, in the company's description, computes several results in parallel. Adding further questions to a single request raises response time only slightly. At those speeds you can afford to run the model over every draft an AI assistant produces before it is sent — checking whether the draft contradicts the earlier correspondence, or promises a refund that is not registered on the customer's account. TypeSafe documents exactly those checks in its pipeline examples, alongside uses in sales and service: detecting intent to buy, sorting tickets by topic, judging the moment a human has to take over.
Source: the-decoder.com
The differentiation problem is the obvious one, and TypeSafe does not dodge it. Ordinary language models already return predefined categories and structured data; OpenAI ships Structured Outputs for precisely this. A constrained response format is not a moat. To matter, Jev has to be not merely faster or cheaper but accurate enough — and that is where the published material thins out. TypeSafe benchmarked four pipelines it had built itself and used other AI models' answers as the reference point rather than independently verified correct solutions. GPT-6 Astra was not in the comparison.
The headline promise deserves the same precision. TypeSafe's claim that Jev cannot hallucinate is a statement about the shape of the answer, not its truth: the model will never emit an option outside the set it was given. It can still pick the wrong option from inside that set. A classifier that returns a delivery label for a billing complaint has not hallucinated by this definition, and the customer is served just as badly. What is being sold as reliability is schema validity, which is a solved engineering problem, and the two are not the same thing.
What looks genuinely new here is the price structure rather than the intelligence. Free output and sub-second latency change what a company can afford to run in the background: not one expensive judgement per conversation but hundreds of cheap ones per turn, on every draft, every routing decision, every escalation check. That is a real shift in the unit economics of a support stack, and it is a smaller, more defensible claim than the one on the label.
The number absent from all of it is accuracy against ground truth. TypeSafe has published latency and price, and has measured itself against other models' opinions. Companies on the waitlist will have to establish on their own data whether a fast, free, structurally valid answer is also the correct one — and by the time they find out, those verdicts will already be feeding rules that act without asking anyone.