What Jev is

Jev is a new AI model category for decisions. An AI workflow is built from two kinds of steps: decisions and actions. Today, the same general-purpose LLM usually does both, so a decision (route this ticket, score this risk, flag this transaction) comes back as a written sentence that then has to be read back into a fixed set of options before an action can follow.

TypeSafe AI (San Francisco) released Jev in early access on September 15, 2026 (TypeSafe’s launch post). It is the first “System One” model (named for the fast, intuitive judgment Kahneman calls System 1 in Thinking, Fast and Slow, as opposed to slower deliberate reasoning). It takes application state plus a set of predefined questions and returns typed decisions: choices, scores, probabilities, confidence values. Unlike a standard large language model, it does not generate prose, hold a conversation, or write code.

Why it matters

Jev is built on the premise that using a general-purpose LLM for automated decisions is slower and more expensive than the decision itself requires. As agentic workflows grow more complex and more widely adopted, TypeSafe’s bet is that this class of decision routing and judgment calls could come to make up a majority of AI inference spend.

TypeSafe reports up to 193.6x faster and 444.6x cheaper than a control large model on workflow evaluations, with 70–500ms latency and $0.042 per million input tokens (output free). If those figures hold up in production, the economics of running high-volume decision steps inside agent workflows change materially.

Jev is a meaningful step toward automation with faster and cheaper agentic workflows built from specialised models. Instead of a single general-purpose LLM handling routing, scoring, and generation alike, the workflow orchestrates the right model for each step.

Considerations

  • Early access, one week old. Released September 15, 2026. Performance figures are vendor-reported, not independently benchmarked or confirmed in production deployments.
  • Narrow scope by design. Jev only classifies and scores against predefined questions. It does not replace a general LLM; it is purpose built for automated decisions within agent workflows.
  • Text and structured data only. Jev does not support image, audio, or video input.
  • No independent latency or accuracy testing found yet. Treat the 193.6x/444.6x figures as directional until third-party benchmarks appear.
  • Proprietary model. TypeSafe has not yet published a paper, architecture details, or model weights.

Recommendation

If you’re already building AI workflows: learn this model class and design with the decision/action split in mind. Treat it as a default design question. Route classification, scoring, and gating steps to a lightweight specialised model, and reserve LLMs for steps that need actual language generation. Don’t commit to Jev itself yet, it’s a week-old early access product from an unproven vendor, but structure new workflows so a decision model can slot in once one is proven.

If you’re not yet building AI workflows: no action is needed now, but know this category exists.

At Foremost, we are testing Jev against a few of our own decision-heavy workflows right now. Curious? Talk to us!

References