Most conversations about AI still assume the output is prose. You ask a question, you get a paragraph. But a surprising share of the decisions that actually run inside production software do not look like that at all. A support ticket needs a department. A lead needs a quality band. A refund request needs a policy verdict. An autonomous agent about to click “send” needs a fast, defensible check on whether the action is safe. None of those are writing tasks — they are classification and judgment tasks dressed up as language.
Jev is TypeSafe AI’s hosted decision model, built specifically for that narrower and considerably less glamorous job. Sometimes referred to as TypeSafe Jev or Jev AI, it takes application state plus a typed question and returns a typed decision rather than generated text. Understanding what that means in practice — and where the model is genuinely strong versus where it needs a careful evaluation — is the difference between a clean architectural fit and an expensive mismatch.
What Jev Is: Typed Questions In, Typed Decisions Out
Strip away the positioning and Jev is a hosted endpoint that accepts two inputs: application state and a typed question. The state can be almost anything you already hold as structured data — a support ticket with its subject line and body, a JSON record from your CRM, a chunk of an order history, a transcript excerpt. You pair that state with a question whose answer shape you declare up front, and the model responds in that shape.
There are three answer types, and each maps to a different family of business decisions. The first is a choice: pick exactly one option from a defined list of up to 255, returned with probabilities and a confidence value. That is the workhorse for routing, tagging, and intent classification, where you need a label and you also want to know how sure the model was before you trust it. The second is a score: place the input on an ordered scale of two to ten levels. That fits urgency ranking, risk tiers, and lead quality bands, where the ordering carries meaning and a numeric level is more useful than a category. The third is a yes/no probability, which TypeSafe refers to as the noul type — a single likelihood you can threshold against your own policy line.
The practical consequence of typed output is that the integration work disappears. There is no parsing step, no regex to catch the model when it adds a courteous sentence before the answer, no retry loop when the format drifts. The response is a number and a label, which means it drops straight into an if statement, a queue assignment, a database column, or a policy engine.
TypeSafe describes Jev as a “System One” model — fast, reflexive, pattern-driven — rather than a deliberative reasoner. That framing is useful because it sets expectations correctly. Jev is not trying to write your customer reply or reason through a multi-step plan. It is trying to make one well-shaped judgment about a piece of state, cheaply, and hand it back. The current release is Jev 1.13, which accepts text and JSON input only (no images), carries a 64k token context window, and ships with official Python and JavaScript SDKs. Documentation and an independent guide to jev and its alternatives are maintained outside TypeSafe’s own site.
Where Jev Fits in Production: Routing, Risk, and Guardrails
The clearest way to understand Jev is through the decisions teams actually wire it into. Ticket routing is the canonical example: take the subject and body of an incoming support message, ask a choice question over your department list, and use the returned probabilities to auto-assign anything above a confidence threshold while sending the ambiguous remainder to a human triage queue. Because the output includes probabilities rather than just a winner, the threshold becomes a tunable dial between automation rate and error rate — a distinction that matters enormously in operations.
Intent classification follows the same pattern with a different taxonomy: billing question, cancellation, feature request, bug report, or one of hundreds of narrower intents. Urgency and risk scoring use the score type instead, placing a conversation on an ordered scale so that a queue can be sorted rather than merely bucketed. Lead quality assessment applies the same machinery to inbound form data and enrichment records.
The most architecturally interesting use case, though, is guardrails. When an AI agent is about to take an action — issue a refund, send an external email, modify a record — a cheap decision model can sit in front of that action and answer a noul question: does this request fall within policy? Is this output safe to deliver? Because the guardrail is a separate, small, typed call rather than another generative pass, it is fast and its verdict is unambiguous. Teams commonly run a cheap decision layer in front of an expensive generative layer, letting Jev absorb the routing and screening volume so the larger model only handles the cases that genuinely need language generation.
Cost and throughput shape those designs directly. TypeSafe’s published pricing is $0.042 per 1M input tokens with output free, which means spend tracks input volume and nothing else — straightforward to model per decision. Throughput is documented at 80 requests a second per key, so high-volume pipelines need to think about queueing or key distribution from the start. TypeSafe also states that customer requests are not used for training, a point that frequently decides whether a regulated team can adopt the tool at all, and there is no listed free tier, so prototype budgets need to be approved rather than assumed.
How to Evaluate Jev: Pricing, Benchmarks, and the Alternatives
Decision models are easy to benchmark badly. A single accuracy number on a public dataset tells you very little about whether a model will classify your tickets, with your taxonomy, at your confidence threshold. That said, published comparisons are a reasonable starting point for narrowing the field, provided you read them as directional signals rather than verdicts.
Preliminary October 2026 benchmarks against DecisionNode, a comparable typed-decision model from Bynn Intelligence, Inc., illustrate the trade-off nicely. Across a mean of 15 text suites, Jev scored 0.803 against DecisionNode’s 0.816 — close enough that the gap is unlikely to decide most production use cases on its own. The divergence appears in the tails. On knowledge questions, Jev leads clearly, posting 0.820 on MMLU-Pro versus 0.698. On prompt injection resistance, the picture reverses: 0.600 for Jev against 0.764 for DecisionNode. For anyone planning to use a decision model as a security or policy guardrail, that second number deserves serious weight, because guardrails are precisely where adversarial input arrives.
Capability boundaries matter as much as scores. Jev accepts text and JSON only, with no image input, which is fine for ticket and record data but rules it out for anything involving scanned documents, screenshots, or video frames. DecisionNode reads images and video frames and adds number, point, and box answer types alongside the choice, score, and probability types — relevant if your decisions involve spatial or numeric extraction. DecisionNode also offers batch discounts, WebSocket streaming sessions, and a free tier, none of which Jev lists; Jev’s counterweights are its knowledge-task strength, its very low input-token pricing, and a clean, narrow answer-type surface that is hard to misuse.
The productive evaluation method is to build a set of a few hundred real state-and-question pairs from your own systems, with the correct answers already recorded from human decisions. Run both models across that set, but score them the way you will actually deploy them: measure accuracy at your chosen confidence threshold, not at the model’s top-1 pick, and track how often each model abstains or returns low confidence. Then price the whole thing out at realistic monthly volume using input tokens per request. A model that is marginally less accurate but far cheaper per decision, or one that is slightly more expensive but meaningfully more robust to injected instructions, will win in different organizations — and only your own evaluation set will tell you which organization you are.
Oslo marine-biologist turned Cape Town surf-science writer. Ingrid decodes wave dynamics, deep-sea mining debates, and Scandinavian minimalism hacks. She shapes her own surfboards from algae foam and forages seaweed for miso soup.