Most AI models in production are asked to write. Jev is asked to decide. It is the hosted decision model from TypeSafe AI — commonly referred to as TypeSafe Jev or Jev AI — and its design goal is considerably narrower, and more practical, than that of a general-purpose chat model: take application state, such as a support ticket or a JSON record, pair it with typed questions, and return typed decisions that software can act on immediately.
What Jev Is: A Decision Model, Not a Chat Model
The distinction between generating text and returning a decision matters more than it first appears. When a language model produces prose, the surrounding application has to parse it, validate it against a schema, handle malformed output, and infer how confident the model actually was. Jev removes that layer by constraining the answer space itself. Rather than free-form output, it returns one of three structured answer types, each designed for a specific class of business rule.
The first is a choice: a single option selected from as many as 255 candidates, returned alongside probabilities and a confidence value. This is the workhorse type for classification problems — picking a queue, a category, a product line, or a next action from a controlled vocabulary. Because the options are supplied by the caller, the model cannot invent a label that does not exist in the schema.
The second is a score: a level on a scale of two to ten ordered levels. Ordered levels are a natural fit for anything graded rather than categorized — urgency tiers, risk bands, lead quality tiers, or severity ratings where the difference between level 3 and level 4 has a defined operational meaning.
The third is the noul type, a yes/no probability. It answers binary questions with a number rather than a word, which lets downstream logic apply its own threshold: approve automatically above 0.95, route to a human between 0.4 and 0.95, reject below 0.4. That thresholding behaviour is where much of the practical value sits, because it turns a model output into a policy decision with explicit, tunable boundaries.
TypeSafe describes Jev as a “System One” model — a reference to fast, instinctive, single-pass judgment rather than deliberative step-by-step reasoning. The framing sets honest expectations. Jev is built for high-volume classification and scoring inside a pipeline, not for open-ended analysis or long-form synthesis. The current release is Jev 1.13. It accepts text and JSON input only, with no image support, and offers a 64k token context window — enough to pass a full ticket thread, a structured record, and evaluation instructions in a single request. Official SDKs exist for Python and JavaScript, each API key is rate-limited to 80 requests a second, and the model does not train on customer requests, a detail that matters to teams working under privacy or regulatory constraints.
Where Jev Fits in Production: Routing, Risk, and Guardrails
The clearest way to understand Jev is through the jobs it is built to do. Its documented use cases cluster around decisions that happen constantly, must be consistent, and are expensive to get wrong: routing, intent classification, urgency and risk scoring, lead quality, policy checks, and guardrails that sit in front of an AI agent before it takes an action.
Consider a support platform receiving several thousand tickets a day. A choice question with a fixed set of queue labels — billing, onboarding, bug report, cancellation, account access — converts each ticket into a routing decision with a confidence score attached. Tickets below a confidence threshold can be diverted to manual triage, so the model handles the easy majority while uncertainty is handled by people rather than guessed at.
Urgency works well as a score question. A ticket describing a total outage and a ticket asking about invoice formatting differ enormously in operational weight, and a model that returns an ordered level rather than a free-text sentiment label gives the workflow something concrete to act on — escalating levels 8 to 10 to an on-call queue, for instance, while scheduling levels 1 to 3 for next-day handling.
Revenue teams apply the same mechanics to lead quality. Given a form submission or CRM record as JSON, a score question ranks the lead, and a noul question answers a narrower binary such as whether the enquiry matches the ideal customer profile. Sales operations can then tune thresholds without retraining anything, adjusting the cut-off rather than the model.
The guardrail use case is arguably the most consequential. Before an autonomous agent issues a refund, changes a subscription, or sends an outbound message, a Jev call can check the proposed action against policy and return a yes/no probability. If the probability that the action violates policy crosses a set boundary, the action is held. This creates a deterministic checkpoint around a probabilistic agent — a pattern that is far easier to audit than asking the agent itself whether it behaved correctly.
Image-heavy workflows are the notable gap. Because Jev reads text and JSON only, tasks that depend on screenshots, scanned documents, product photos, or video frames sit outside its scope and require either a separate perception step or a different model entirely.
Cost, Limits, and How to Evaluate Jev Against Alternatives
Pricing for Jev is unusually shaped and worth understanding before scoping a project. Per TypeSafe’s documentation, checked in October 2026, the model costs $0.042 per 1 million input tokens, with output free. That structure rewards workloads that send substantial context in and receive small structured answers out — which is precisely the shape of a decision call. There is no listed free tier, so any pilot requires a paid key from the start, and budgeting should account for evaluation runs as well as production traffic.
Practical limits are similarly clear-cut. The 64k token context window comfortably accommodates long ticket histories and multi-field JSON records. The 80 requests per second per key ceiling is generous for most internal workflows but becomes a planning constraint for consumer-scale, burst-heavy applications, where queueing or multiple keys may be needed. Official Python and JavaScript SDKs reduce integration effort, and the commitment not to train on customer requests makes the model easier to justify in environments where data handling is reviewed by legal or security teams.
Because the model is narrow by design, the sensible question is not whether it is generally capable but whether it is better than the alternatives at the specific decision types a team actually runs. For that, a structured reference is useful: this independent jev guide sets out the pricing, limits, and comparative benchmarks in one place, alongside profiles of competing models. One measured alternative, DecisionNode from Bynn Intelligence, also answers typed questions with probabilities, but additionally reads images and video frames, adds number, point, and box answer types, and offers batch discounts, WebSocket streaming sessions, and a free tier — differences that matter if multimodal input or high-volume batching is on the roadmap.
Preliminary October 2026 benchmarks published for comparison put the two models close on aggregate text performance: a mean of 0.803 for Jev against 0.816 for DecisionNode’s mixture-of-experts configuration across 15 text suites. Jev leads on knowledge-heavy questions, scoring 0.820 on MMLU-Pro against 0.698. DecisionNode leads on prompt injection resistance, at 0.764 against 0.600 — a gap worth weighing carefully for any deployment where untrusted user input reaches the model. Aggregate averages hide a great deal, however, and a decision model’s value is determined by accuracy on one’s own categories, thresholds, and edge cases. The most reliable path is a private evaluation: assemble a few hundred labelled examples from real traffic, run both models against identical typed questions, and compare confidence calibration as well as raw accuracy before committing production traffic to either.
Madrid linguist teaching in Seoul’s K-startup campus. Sara dissects multilingual branding, kimchi microbiomes, and mindful note-taking with fountain pens. She runs a weekend book-exchange café where tapas meet tteokbokki.