Most teams that bring a language model into a production workflow hit the same wall: the model returns prose, and the surrounding software needs a value it can branch on. Jev, the hosted decision model from TypeSafe AI, is built around that gap. Instead of generating a paragraph for a downstream system to parse, it accepts application state — a support ticket, a JSON record, an order payload, a CRM entry — together with typed questions, and returns typed decisions. TypeSafe describes it as a “System One” model: fast, narrow, and intended to sit in front of the slower reasoning layers of an AI stack.
The current release is Jev 1.13, usually referred to as TypeSafe Jev or Jev AI. It is a hosted service rather than a model you deploy yourself, and it is deliberately limited in scope: text and JSON input, structured output, and a per-token price aimed at high-volume classification and routing rather than long-form writing. This guide explains what the model returns, where it fits in production, and how to evaluate it against the alternatives.
What TypeSafe Jev Returns: Choices, Scores, and the Noul Type
The defining feature of Jev is that every request must declare the shape of the answer before the model runs. Three answer types are supported, and each maps cleanly onto a data structure your application already understands.
A choice returns exactly one option from a set you define, which can hold up to 255 candidates. Alongside the selected option, Jev returns a probability distribution across the candidates and a confidence figure. That matters in practice: a routing decision with 0.94 confidence can be applied automatically, while a 0.41 confidence on a four-way split is a signal to escalate to a human reviewer or a larger model. Without the probability attached, both cases would look identical to your code.
A score returns a level on an ordered scale of between 2 and 10 levels. This is the natural fit for anything that is genuinely ordinal — urgency, risk, lead quality, severity, churn likelihood — where forcing the answer into a fixed label set loses information. Because the scale is explicit and bounded, the output can be sorted, thresholded, and aggregated across records without post-processing.
The third type, called a noul, returns a yes/no probability. It is the most useful primitive for gating logic: does this message contain a refund request, does this record satisfy the policy, should this action be blocked pending review. The probability rather than a hard boolean is what makes threshold tuning possible later, when you have collected labelled outcomes and want to move the cut-off.
Two constraints shape how the model is used. Input is text and JSON only — Jev does not accept images — and the context window is 64k tokens, which is generous for tickets, documents, and structured records but not intended for bulk corpus work. The value of the design is that a downstream system never has to parse a natural-language answer, guess at a label, or repair malformed output. The schema is known in advance, the response conforms to it, and the probabilities are exposed rather than hidden.
Where Jev Earns Its Place: Routing, Risk Scoring, and Agent Guardrails
The workloads TypeSafe targets are the ones that sit on the critical path of an application and therefore cannot afford latency variance or unconstrained output. Four patterns come up repeatedly.
Routing and intent classification. A support ticket arrives, and something must decide which queue, team, or workflow owns it. Framed as a choice over the queue list, with probabilities attached, this becomes a single call that returns a queue plus a confidence. Low-confidence tickets route to a triage queue instead of being misassigned, which is usually the more expensive failure.
Urgency and risk scoring. Security alerts, customer complaints, and flagged transactions all need to be ranked before anyone looks at them. Framed as a score on a bounded scale, the output feeds directly into prioritisation logic and SLA timers. The ordinal nature of the answer is what makes the ranking meaningful rather than a reshuffled label set.
Lead quality and qualification. Inbound form fills and enrichment records can be scored against the same criteria every time, with the score driving routing to sales, nurture, or disqualification. Because the model does not train on customer requests, the same input produces a consistent decision regardless of what other tenants are doing.
Policy checks and guardrails before an agent acts. This is where typed decisions become architecturally interesting. When an autonomous agent proposes an action — issue a refund, send an outbound email, modify a record — a decision model can be placed in front of it as a gate. The question is a noul: does this action violate a stated policy. A low yes-probability lets the action proceed; a high one blocks it and explains nothing, because nothing needs explaining — the block is a structured value the orchestration layer already knows how to handle. Placing a cheap, deterministic-shaped model in front of an expensive generative agent also reduces the number of costly calls, since clearly routine inputs never reach the reasoning layer at all.
A useful way to sanity-check the fit is to ask whether the question could be answered by a form. If the desired output is a field with a bounded set of values, a level on a scale, or a boolean with a threshold, Jev is shaped for it. If the desired output is a paragraph, a plan, or an explanation a human will read, it is the wrong tool, and a general chat model remains the better choice.
Pricing, Throughput, and How Jev Compares on Benchmarks
The commercial model is simple and volume-oriented. According to TypeSafe’s documentation as of October 2026, Jev is priced at $0.042 per 1M input tokens with free output. There is no listed free tier, so any evaluation requires a paid key from the start. Throughput is capped at 80 requests a second per key, and official SDKs are available for Python and JavaScript. TypeSafe states that customer requests are not used for training. For teams planning a rollout, the per-key rate limit is often the binding constraint rather than cost — a routing workload spread across several keys scales horizontally, but a single-key integration will hit the ceiling under bursty traffic.
Benchmark comparisons are where the choice gets less obvious. Preliminary October 2026 measurements against DecisionNode, a comparable typed-decision model from Bynn Intelligence, Inc., put the two close on general text work: across a mean of 15 text suites, Jev scored 0.803 against 0.816 for DecisionNode’s MoE configuration. Jev led clearly on knowledge questions, with MMLU-Pro at 0.820 versus 0.698. On prompt injection resistance the gap ran the other way and was wider, at 0.600 for Jev versus 0.764 for DecisionNode.
The functional differences matter as much as the numbers. DecisionNode also answers typed questions with probabilities, but additionally reads images and video frames, supports number, point, and box answer types, offers batch discounts and WebSocket streaming sessions, and provides a free tier. If your records include scanned documents, screenshots, or frames, Jev’s text-and-JSON-only input is a hard boundary rather than a preference.
The practical takeaway is to evaluate on your own data before committing. Build a labelled set of a few hundred real records, run both models against the same questions, and compare agreement, confidence calibration, and failure modes at your chosen threshold. A maintained typesafe jev reference is useful for tracking version changes, pricing updates, and benchmark results as they are published, since the numbers above are preliminary and the model versions move. Treat the published figures as a screening tool and your own evaluation set as the decision.
