AI systems are often asked to do two very different jobs at once: understand a messy situation and commit to a reliable action. Generated text can help with the first job, but production systems need something more constrained for the second. Jev, the hosted decision model from TypeSafe AI, is built around that distinction. Instead of returning free-form prose, Jev takes application state—such as a support ticket or a JSON record—plus typed questions, and returns typed decisions. That design makes it easier to route work, score risk, classify intent, and enforce policy checks before an AI agent takes an action.
From Generated Text to Typed Decisions: The Core of Jev
Jev is often described as “TypeSafe Jev” or “Jev AI,” and its central idea is simple: a decision should have a shape. TypeSafe calls Jev a “System One” model, which signals that it is intended for fast, structured judgments rather than open-ended conversation. The model accepts text and JSON input and supports a 64k token context window. It does not accept images. For teams working with ticketing systems, CRM records, form submissions, or application logs, that input profile covers a large share of operational data.
The output types are where Jev becomes especially useful for production workflows. A choice answer selects one option from up to 255 possibilities and returns probabilities plus a confidence value. A score answer places an item on 2 to 10 ordered levels. A yes/no probability—the “noul” type—expresses uncertainty without forcing an immediate binary. These formats map naturally to routing, intent classification, urgency and risk scoring, lead quality, policy checks, and guardrails before an AI agent takes an action. Instead of parsing generated text and hoping the format holds, a system can consume a typed decision directly.
The current version is Jev 1.13. TypeSafe provides official Python and JavaScript SDKs, which lowers integration effort for teams already building in those languages. The model is available at 80 requests a second per key, and TypeSafe states that customer requests are not used for training. That last point matters for organizations that need to send sensitive ticket text or JSON records through a hosted decision layer. An independent guide to jev and its alternatives can help teams compare these operational details, but the core value proposition is already clear: Jev is designed to convert state plus questions into decisions that software can act on.
Pricing, Throughput, and Performance Trade-Offs in Jev 1.13
Cost and throughput shape how a decision model fits into production. According to TypeSafe’s documentation checked in October 2026, Jev is priced at $0.042 per 1M input tokens with free output. There is no listed free tier. The model supports text and JSON input only, with no image support, and it runs with a 64k token context. Throughput is listed at 80 requests a second per key. These are not abstract numbers: they determine whether a team can place Jev in a synchronous request path, use it for batch enrichment, or reserve it for high-value decisions where the cost per call is justified.
Jev is not the only hosted model in this category. The independent guide compares it with DecisionNode, a measured alternative from Bynn Intelligence, Inc. DecisionNode also answers typed questions with probabilities, but it adds capabilities that Jev does not list: it reads images and video frames, adds number, point, and box answer types, offers batch discounts, WebSocket streaming sessions, and a free tier. For teams whose decision inputs are purely text or JSON, those differences may be irrelevant. For teams that need to evaluate visual evidence or want to experiment without a paid commitment, the comparison becomes more significant.
Preliminary October 2026 benchmarks provide a useful starting point. Across a mean of 15 text suites, Jev scored 0.803 while DecisionNode MoE scored 0.816. On prompt injection, Jev scored 0.600 versus 0.764 for DecisionNode. Jev led on knowledge questions, with MMLU-Pro 0.820 versus 0.698. These are preliminary figures, not final verdicts, but they suggest a trade-off pattern: Jev shows strength on knowledge-oriented questions, while the alternative posts stronger prompt-injection resistance and a slightly higher mean across the tested text suites. Teams should treat benchmarks as screening tools, then test their own ticket text, JSON schemas, choice sets, and ordered scoring rules.
Evaluating Jev for Routing, Risk, and Guardrails in Production
Real-world evaluation should begin with the decisions the system actually needs to make. A support organization might use Jev to route a ticket to the right queue, classify intent from a short customer message, or score urgency on an ordered scale. A revenue team might ask whether a lead is qualified, using a yes/no probability rather than a generated explanation. A policy or safety team might place Jev as a guardrail before an AI agent takes an action, checking whether a request falls inside an allowed category. These scenarios share a common pattern: the input is structured or semi-structured state, the question is typed, and the output must be a decision that downstream software can enforce.
Evaluation should test the answer types directly. For choice decisions, teams should inspect probability distributions and confidence values, not just the selected option. A high-confidence wrong choice is more dangerous than a low-confidence correct one in many workflows. For score decisions, ordered levels from 2 to 10 need calibration: a “7” should mean the same thing across ticket types and JSON records. For yes/no probability, thresholds should be tuned against real operational costs. Prompt-injection tests are also essential, especially because the preliminary benchmark puts Jev at 0.600 on prompt injection versus 0.764 for DecisionNode. A decision layer that sits before an AI agent takes an action must be tested with adversarial inputs, not only clean examples.
Operational constraints belong in the same evaluation. Jev’s 80 requests a second per key, 64k token context, and $0.042 per 1M input tokens with free output define a different cost and scaling profile than a model with a free tier or batch discounts. Teams should estimate peak request volume, average input size, and whether text-only input is sufficient. The absence of image support is a hard boundary for workflows that depend on screenshots, scanned forms, or video frames. The fact that TypeSafe does not train on customer requests is a relevant data-governance detail. Jev’s version—1.13—should also be pinned in evaluation notes so that benchmark results and production behavior remain comparable over time. A structured pilot with typed questions, confidence thresholds, and a measured alternative provides a more reliable path than relying on a single aggregate score.
