Last week, TypeSafe AI introduced Jev, a new category of model it calls a decision model. Unlike an LLM, Jev is not designed to generate prose. It takes unstructured text or JSON and returns a typed decision: a category, score, yes/no answer, extracted value, and an associated confidence score.

TypeSafe describes it as “a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.” The important distinction is that a program can branch directly on the result instead of interpreting generated text.

Why It Is Interesting

Jev is essentially a smart if statement for cases where hand-written rules are too brittle but a full LLM call is too slow or expensive. Its System One API exposes three question types: Choice selects from caller-defined options, Score rates against ordered criteria, and Noul estimates whether a statement is true. Choice and Score return a probability distribution and confidence; Noul returns a probability from 0 to 1.

One request can ask several independent questions against the same state. TypeSafe says those questions are evaluated in parallel and isolation, so an application can separate a broad judgment into atomic questions, then combine their outputs deterministically in code. The API is available through SDKs or POST /v1/systemone; the quickstart shows the request and response shapes.

Useful patterns include:

  • Routing a request to the right workflow or model
  • Scoring relevance, risk, urgency, or quality
  • Extracting structured values from text
  • Applying guardrails before an expensive downstream call
  • Making a bounded decision that application code can immediately act on

TypeSafe says its current Jev release is priced at $0.042 per million input tokens with no output charge. It supports a 64K-token request budget, with a 32K-token limit for the state plus the longest question; inputs are text only, including strings, JSON objects, and arrays of text. The models documentation has the current pricing, context, rate-limit, and version-alias details. These are compelling claims if they hold up in production, especially for high-volume classification and routing workloads.

Because the response is constrained to a caller-provided schema, Jev cannot return a value outside that schema. That is a meaningful reliability property, but it should not be confused with correctness: a schema-valid decision can still be confidently wrong. TypeSafe trains Jev with reinforcement learning for calibrated decisions (RLCD), meaning probabilities are intended to track accuracy across groups of predictions, not guarantee that an individual output is correct. Its AI primer explains that distinction.

Access and Open Alternatives

Direct access to Jev remains limited through a waitlist, which means independent validation of TypeSafe’s frontier-model claims will take time. Vercel AI Gateway provides a way to call typesafe-ai/jev and documents per-request Zero Data Retention and no-training controls. Those controls are useful for experiments with sensitive prompts, but teams should still validate the applicable plan, provider agreement, and data-handling terms before production use.

Jev is a hosted proprietary model, not a package that can currently be deployed directly on AWS or Databricks. Open projects can offer a self-hosted alternative, but they are reproductions of the decision-model interface rather than open-source Jev itself. SemIf, formerly OpenJev, reads option probabilities from an open model; Kev provides Jev-compatible open-weight models; and Laya is another open decision-model approach. The Jev Decision Index tracks Jev and alternatives, but benchmarks should be treated as point-in-time and method-dependent rather than proof of equivalence.

One additional claim in the original draft is that some reference answers used GPT-6 Astra and Fable 5.1. I have not found a primary source confirming that attribution, so it should be treated as unverified rather than evidence about Jev’s training or evaluation process.

Reasons for Caution

Jev is still new and access is limited. Teams working with sensitive data would need to resolve deployment, data-retention, and contractual requirements before experimenting with it. It accepts text and JSON only, so it is not a fit for image understanding, open-ended reasoning, long-form writing, or code generation.

There are also important unanswered questions. TypeSafe has not published an architecture paper, and the model’s training data and decision-calibration process are not public. Confidence scores are useful only if they are calibrated on the workload at hand. TypeSafe’s own documented limitations advise keeping arithmetic, counting, date comparisons, and structural invariants in code; filtering irrelevant state; and using a generative model for generation. Like any model, Jev may also perform unevenly across domains or when presented with a large number of competing options.

Decision models are worth watching because they offer a potentially cheaper, faster building block between deterministic rules and general-purpose LLMs. The practical test will be whether their decisions remain accurate, calibrated, and robust outside vendor benchmarks.

Further Reading