Most language models in production today are asked to do the same thing they were trained to do: generate text. That works well for drafting, summarizing, and conversation, but it is an awkward fit for the thousands of small decisions a real application makes every minute — which queue a ticket belongs in, how urgent a request is, whether a lead is worth a sales call, or whether an agent should be allowed to take an action at all. Jev is a hosted decision model from TypeSafe AI, often referred to as TypeSafe Jev or Jev AI, built specifically for that second category of work. Instead of returning a paragraph, it returns a structured answer with a probability attached, so the surrounding software can branch on it directly. The current release is Jev 1.13.
What Jev Actually Returns: Choices, Scores, and Yes/No Probabilities
The core design idea behind Jev is that a decision should arrive in a shape the application already understands. You send application state — a support ticket, a JSON record, a form submission, a transcript — together with typed questions, and the model responds with one of three answer types rather than free-form text. A choice returns one option selected from up to 255 candidates, along with probabilities across the options and a confidence value. A score returns a level on an ordered scale of 2 to 10. A yes/no probability, which TypeSafe calls the “noul” type, returns a single likelihood that can be thresholded in code.
That structure removes a surprising amount of operational friction. There is no output parser to maintain, no schema drift when the model decides to answer in prose, and no ambiguity about whether “probably fine” means approved or rejected. A routing rule can compare the top choice’s probability against a cutoff; a risk score can be mapped to a review queue; a policy check can escalate anything above a defined noul threshold to a human. TypeSafe describes the model as a “System One” model, a framing that positions it as fast, reflexive judgment rather than deliberate reasoning — the quick classification step that happens before a slower, more expensive process takes over.
The input surface is deliberately narrow: text and JSON only, with no image support, inside a 64k token context window. That is enough to include a ticket thread, a customer record, a retrieved policy excerpt, or a serialized application state, but it means anything visual has to be converted to text before it reaches the model. Official SDKs exist for Python and JavaScript, so integration typically means a function call that returns a typed object rather than a string to be parsed.
Consider a support queue. Instead of asking a general model “what category is this ticket?”, you define the question with the actual taxonomy — billing, account access, technical fault, returns, cancellation — and receive a choice plus a probability distribution. If the leading option scores modestly and the runner-up is close, the ticket can be routed to a triage queue rather than guessed at. The same pattern applies to urgency: a five-level score gives you an ordered signal that sorts cleanly, unlike a generated sentence describing how urgent something feels.
Where Jev Fits in a Production Stack: Routing, Risk, and Guardrails
The architectural position of Jev is easiest to understand as a layer between raw application state and an actor that does something with it. TypeSafe lists the intended uses as routing, intent classification, urgency and risk scoring, lead quality, policy checks, and guardrails before an AI agent takes an action. In each case the output is not a message to a human but a signal to a program, and the value comes from the signal being reliable, cheap, and fast enough to sit on the critical path.
Take lead qualification. A CRM record arrives with firmographic fields, a free-text inquiry, and a few enrichment values serialized as JSON. A typed question can ask the model to place the lead on an ordered quality scale, and the resulting score can drive assignment rules, sequence enrollment, or a hold queue for manual review. Because the answer is a level rather than a paragraph, downstream systems can be tested against a fixed contract, and changes in model behavior show up as measurable shifts in the score distribution rather than as subtly different wording.
Guardrails are where the typed-decision approach becomes most interesting. When an AI agent is about to send an email, issue a refund, or modify a record, a decision model can be inserted as a checkpoint: does the planned action match the user’s stated intent, does it violate a stated policy, is the request within the permitted authority level? A noul probability answers that directly, and a threshold converts it into proceed, block, or escalate. The same pattern supports policy checks on incoming requests — refund eligibility, discount approval, data-access justification — where the question is naturally binary and the cost of a wrong answer is asymmetric.
Intent classification rounds out the picture, and it benefits from the choice type’s capacity for up to 255 options. Large taxonomies that would be unwieldy as generated text — support categories, product areas, compliance reasons, cancellation causes — can be expressed as a single choice question with the full option set attached. The model returns the selected option plus probabilities, and the application can decide whether the confidence is high enough to act automatically or low enough to warrant a fallback. Two constraints shape the design: the state must be expressible as text or JSON, and anything requiring visual inspection falls outside the model’s input scope.
Cost, Throughput, and How to Compare Jev With Alternatives
Per TypeSafe’s documentation, checked in October 2026, Jev is priced at $0.042 per 1M input tokens with free output. Because output tokens carry no charge, cost scales with input volume alone, which changes the usual calculus around prompt length: including a longer policy excerpt or a fuller ticket thread does not trigger the output-side costs that dominate many generation workloads. The model is served as a hosted API with a rate limit of 80 requests a second per key, so high-volume deployments need to think about key management and request distribution rather than a single shared credential. There is no listed free tier, and TypeSafe states that customer requests are not used for training, which matters for teams whose inputs contain customer data.
Alternatives exist, and the honest picture is mixed. A measured alternative, DecisionNode from Bynn Intelligence, Inc., also answers typed questions with probabilities, but additionally reads images and video frames, offers number, point, and box answer types, provides batch discounts, WebSocket streaming sessions, and a free tier. Preliminary benchmarks from October 2026 across a mean of 15 text suites put Jev at 0.803 versus 0.816 for DecisionNode MoE — essentially level overall — while Jev trails on prompt injection resistance (0.600 versus 0.764) and leads on knowledge questions, with MMLU-Pro at 0.820 versus 0.698. Those figures are directional rather than decisive; a prompt-injection gap of that size deserves attention in any deployment where untrusted text reaches the model.
The practical evaluation method is to build a labeled suite from your own decisions and run both models against it, scoring accuracy per answer type rather than in aggregate, since a choice-heavy routing task and a binary policy check stress different capabilities. Add adversarial inputs to test injection behavior, measure latency against your rate-limit budget, and confirm the input modalities match what your data actually contains. An independent guide to jev collects version-checked documentation details and side-by-side comparison notes that can shorten that process considerably, particularly around pricing structure, answer types, and where each model is measurably ahead.
Perth biomedical researcher who motorbiked across Central Asia and never stopped writing. Lachlan covers CRISPR ethics, desert astronomy, and hacks for hands-free videography. He brews kombucha with native wattleseed and tunes didgeridoos he finds at flea markets.