TypeSafe AI has opened early access to Jev, an AI model that returns decisions software can use directly: choices, scores and probabilities instead of prose. The company emerged from stealth on September 15, 2026, with a $40 million seed round led by DCVC. Jev’s listed price is $0.042 per million input tokens, a measure of the text sent to the model.
The practical possibility is a support request reaching the right team without waiting for a model to compose an answer. Jev supplies the judgment; the surrounding software decides what happens next. TypeSafe reports response times of 70 to 500 milliseconds and says it is admitting developers from its waitlist.
That speed is a company measurement, not a service guarantee. TypeSafe says its published evaluations generally ran from West Coast laptops near its West Coast service, and its short-input demonstration favors Jev. The range does not establish performance under heavy traffic or from other regions. Its documented price includes free outputs, though TypeSafe acknowledges that the price’s long-term sustainability remains unproven.
According to the API documentation, developers supply context, such as a customer’s message and account information, alongside narrow questions. Choice selects from listed options, Score evaluates a rubric, and Noul estimates the probability that a statement is true. Questions in one request are evaluated independently and in parallel. Code combines the answers, and later steps can depend on earlier results. Developers still have to define those questions and connect the answers to appropriate actions; the launch does not establish chatbot replacement or staffing reductions.
TypeSafe’s claim that Jev “can’t hallucinate” needs a narrower reading. The company says responses always match the predefined output structure, and its zero type-error figure follows from that guarantee rather than an empirical test. That addresses malformed or out-of-range answers, not the truth of a decision. Its own Jev 1.13 failure notes describe unreliable counting and date comparisons, distraction by irrelevant information and susceptibility to adversarial text.
Imagine a customer writes, “I was charged twice”: software could give Jev the message and relevant account details, then ask it to choose billing, delivery, account or other, score urgency and estimate whether the customer wants a refund.
Code could route or escalate the ticket using thresholds tested on real examples, while applying transaction-record and payment-eligibility rules; a person or text-generating model would write the reply. This example was not run, and a permitted label such as “delivery” could still be wrong, even with high reported confidence.
Confidence needs its own test. TypeSafe’s documentation describes Choice and Score confidence as a summary of the returned probability distribution: how strongly the model favors an answer. A confidence score of 0.9 cannot simply be read as 90% accuracy. Independent calibration testing should establish whether predicted probabilities track actual outcomes on unfamiliar cases, including those software accepts automatically.
The company’s workflow evaluations offer an initial comparison of models across customer service, invoice processing, security incidents and review of AI-agent activity. TypeSafe describes its reference answers as averages of GPT-6 Astra and Claude Fable 5.1 probabilities at high reasoning settings. Agreement with those references is not independently verified accuracy: the reference models can share mistakes.
The tests assume the workflow code is correct. TypeSafe’s capabilities team built the workflows, while competing models use their providers’ default reasoning settings and software requiring probability outputs. TypeSafe says this approach tends to be slower and more expensive than requesting decisions without probabilities. These choices limit how broadly the comparisons apply; the company itself says its advertised speed and cost multiples probably sit toward the high end of real-world gains.
For support teams, Jev’s value will turn on decisions checked against real tickets and the work left after an error. Production records of misroutes, escalations and correction time, alongside response speed, would show whether cheap judgments actually mean less rework.