TypeSafe AI launches Jev scoring model

TypeSafe AI has introduced a model called Jev that is designed to score predefined options rather than generate free-form text, a former OpenAI researcher says the approach can deliver narrow judgments inside software in tens to hundreds of milliseconds.

According to The Decoder, Jev is aimed at the kind of background decisions that companies already try to automate: classifying customer messages, estimating intent, or deciding when a human should step in. Developers define the questions and the allowed answers, and Jev returns labels and probabilities. In the example described, customer service software routes “payment issue” messages to accounting and flags likely refund requests for closer review. The model’s value proposition is not creativity but constraint: it cannot wander outside the permitted output structure, which is attractive to businesses that have been burned by chat-style systems producing plausible but unusable prose.

The pitch also reflects a shift in how AI is being operationalised. Many firms do not need a chatbot; they need a fast, cheap classifier that can be called repeatedly as part of a workflow, including as a check on other models. The Decoder reports TypeSafe claims Jev responds in roughly 70 to 500 milliseconds by computing several outputs in parallel and that adding more questions in the same call barely increases latency. If that holds under real traffic, it makes “judge before you act” architectures more feasible: a front-line assistant drafts a response, while a second model scores consistency, policy compliance, or whether a refund claim looks like a refund claim.

But the constraints are not the same as correctness. Jev will not invent a new category, yet it can still pick the wrong one, and the downstream software will still execute whatever rules the company has attached to that label. The Decoder notes that ordinary language models can also be forced into preset categories using structured outputs, including offerings from OpenAI, making differentiation a matter of speed, price, and accuracy under independent evaluation. TypeSafe’s published tests, as described by The Decoder, compare workflows the company built itself and use other AI models as reference points rather than externally verified ground truth.

TypeSafe lists pricing at $0.042 per million input tokens and charges nothing for outputs, a structure that encourages frequent small calls. That pricing model, combined with the promise of predictable outputs, is aimed squarely at businesses that want automation without the reputational risk of a chatbot improvising in public.

Jev is marketed as a model that cannot hallucinate outside the allowed format. The harder question for buyers will be how often it makes the wrong choice inside that format, and what their software does when it does.