TypeSafe Launches Jev, a Low-Cost Quick Choice AI

TypeSafe AI unveils Jev, a model that makes judgments and probabilities instead of generating text. The company, co-founded by Diogo Almeida, a former OpenAI employee and co-author of InstructGPT, announces response times of 70 to 500 ms and a price of $0.042 per million input tokens, with free outputs. Access begins via a waitlist, and the published tests come with limitations.

No guaranteed hallucinations on form, not on substance

TypeSafe presents Jev as a model that cannot hallucinate. This promise pertains only to the structure of the response: the model remains limited to predefined options, but it can still select an incorrect option in terms of substance. Jev is designed to perform numerous discrete judgments in the background at a lower cost, but the reliability of these decisions will need to be assessed by each company according to its needs. The published performance tests have limitations: TypeSafe compares four workflows it has created and uses responses from other models as a reference, without independent validation and without including GPT-6 Astra. Furthermore, traditional language models can already produce categories and data structures, and OpenAI offers Structured Outputs. Therefore, a structured format alone is not enough to distinguish Jev; it must also provide advantageous speed or cost without sacrificing quality.

Announced pricing and limited access at launch

TypeSafe announces a price of $0.042 per million input tokens and specifies that outputs are not charged. Access to Jev for developers begins with a waitlist. The model aims to make many automatic judgments in the background more economical.

How Jev integrates into a business workflow

Developers define the questions and allowed responses, then link their software to Jev. For each received message, the text is sent to the model, which returns labels and probabilities, such as "payment issue" and the likelihood that a refund is desired. The software then applies fixed rules, such as directing payment issues to accounting or flagging refund requests. If the label is uncertain, a staff member can intervene. Jev provides the assessment, but the final decision depends on the software and its rules.

Under the announced second and use cases for control

TypeSafe claims that Jev responds in 70 to 500 milliseconds, which would be several times faster than the fastest language models. According to the company, this performance is due to the absence of step-by-step text generation and the parallel computation of multiple outputs, with the addition of questions in the same call only slightly increasing latency. This speed could allow for checks before each response from an AI assistant, such as detecting a contradiction with the ongoing conversation or an unrecorded mention of a refund. TypeSafe describes these uses in its workflow examples and also cites applications in sales and customer service, such as detecting purchase intent, thematic sorting of requests, or identifying situations requiring human intervention.