TypeSafe launched a new artificial intelligence model last week called Jev. The technology behaves differently from standard frontier models by restricting its output instead of using a chatty approach. Jev acts as a classifier that returns predefined typed decisions, probability distributions, and confidence levels.
In the week since its release, developers have tested the service through various prototype applications. FPV Ventures partner Nikunj Kothari launched a site named Jevable to collect these projects from X. Applications range from spreadsheet urgency raters to virtual clothing try-on tools that cost $0.0011 per decision.
Coder curiosity has even led developers to use Jev to emulate a CPU. This led to the creation of the term JevOps to describe running all code on Jev-based virtualization. Other projects include gaming hacks for titles like Doom and chess, where Jev lost to an open-weight model but ran much more cheaply.
TypeSafe describes Jev as a System One model designed for fast and structured decisions. Users submit data to an API and ask three specific types of questions. These primitives are called a Choice, a Score, or a Noul, which is a probability score of truthfulness between zero and one.
Using Jev requires manual configuration because users must define schemas and candidate options ahead of time. TypeSafe recommends asking targeted questions like whether a message conveys urgency instead of open-ended analysis requests. This structured response allows Jev to return answers in as little as 150 milliseconds.
The service costs $0.042 per million input tokens with no charge for output tokens. TypeSafe technical staff member Eugene Shvarts noted in a GitHub README that the model enables many affordable semantic judgments within application response-time budgets. Potential workloads include job recruiting and scientific paper screening.
AI personality Mo Bitar questioned the actual intelligence and accuracy of the model in a video review. Engineer Archer Hume probed Jev with 10,000 API calls and surmised that it calculates decision probabilities directly from internal representations without predicting text.
Industry figures have praised the approach as a shift toward smarter systems rather than just smarter frontier models. Swamp Club CEO Adam Jacob stated that correct and repeatable results are necessary. Anthropic researcher Andrej Karpathy wrote on X that Jev addresses latent demand for a single-token LLM with low latency.
Developers continue to test the API service to determine its limits and best use cases for production environments.
