Jev is a structured decision model from TypeSafe AI: instead of writing an answer, it picks one — a yes/no with a probability, an option from a defined set, or a score on a scale — and returns it as a typed value your code can consume directly.
Jev is the first model in TypeSafe's System One class. Where a conversational model writes an answer, Jev makes a decision: it returns a single typed value that conforms to a schema you define, so your application can consume it directly without parsing prose.
According to TypeSafe, it handles three kinds of question — a yes/no answer with a probability attached, a choice among predefined options, or a score against a scale. Selections can span up to 255 discrete options.
On AI/ML API it is served as typesafe/jev, which currently resolves to Jev 1.13. The model id stays the same as TypeSafe ships new versions.
Typed output by construction
TypeSafe states that schema matching is guaranteed and that type errors are mathematically impossible, because the output space is constrained to the schema rather than sampled as free text.
Decision latency, not generation latency
With no autoregressive decoding, TypeSafe reports end-to-end responses between 70 and 500 milliseconds.
No output bill
Output is unmetered. This is structural rather than promotional: there are no output tokens to count when nothing is decoded.
TypeSafe's headline comparisons — up to 193.6× faster and roughly 444× cheaper than conversational models — are vendor-reported and self-scored. They have not been independently reproduced, and several outlets covering the launch flagged exactly that. Treat them as the vendor's own claims until third-party numbers exist.
The "cannot hallucinate" claim applies to the shape of the output, not its correctness: Jev will always return a valid value from the allowed set, but that does not guarantee the decision is right.
TypeSafe has said its own rate may be subsidised and is expected to fall over time.
Routing and branching
Deciding which path a request should take, where a typed answer is all the calling code needs.
Classification at scale
Scoring or bucketing high volumes of input where a free-text answer would only have to be parsed back into a value.
Latency-bound paths
Places where a conversational round trip is too slow to sit in the request path.
Sources: https://openrouter.ai/~typesafe/jev-latest · https://www.artificialintelligence-news.com/news/chatgpt-pioneer-launches-jev-model-for-programmatic-logic/ · https://www.testingcatalog.com/icymtypesafe-ai-launches-jev-for-structured-ai-decisions/
Measured on the AI/ML API gateway across live traffic. Figures are medians unless stated. Last updated 2026-10-12.
Jev 1.13 is documented as suited for routing, where it can make a fast, structured decision about which path an application should take. It is also documented for classification, including picking labels or other typed outcomes from input text or JSON state.
Another documented use case is moderation, where the model can evaluate content and return a decision for software to act on. It is also described for LLM-as-a-judge scoring and rubric-based evaluation, where it scores input against criteria and returns values code can branch on.
import requests r = requests.post( "https://api.aimlapi.com/v1/decisions", headers={"Authorization": "Bearer " + AIMLAPI_KEY}, json={ "model": "typesafe/jev", "state": "Help! My payments have been failing for 3 days.", "questions": { "urgent": { "type": "noul", "instructions": "Is this an urgent support request?" } } }, ) print(r.json())
const r = await fetch("https://api.aimlapi.com/v1/decisions", { method: "POST", headers: { Authorization: `Bearer ${process.env.AIMLAPI_KEY}`, "Content-Type": "application/json", }, body: JSON.stringify({ "model": "typesafe/jev", "state": "Help! My payments have been failing for 3 days.", "questions": { "urgent": { "type": "noul", "instructions": "Is this an urgent support request?" } } }), }); console.log(await r.json());
curl -X POST https://api.aimlapi.com/v1/decisions \ -H "Authorization: Bearer $AIMLAPI_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"typesafe/jev","state":"Help! My payments have been failing for 3 days.","questions":{"urgent":{"type":"noul","instructions":"Is this an urgent support request?"}}}'
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Input | |
| Output | |
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
Jev 1.13 This page | Typed decisions inside software | |||
| High-volume, cost-efficient agentic tasks | ||||
| Reasoning + agents |
Yes. Jev has been live on AI/ML API since 17 September 2026 under the model id typesafe/jev, served through the POST /v1/decisions endpoint. Input is billed at $0.0577668 per 1M tokens; output is free.
System One is TypeSafe's name for a class of model that returns a decision rather than prose. Jev is the first of them: instead of generating text, it emits a single typed value matching a schema you define, so your application can consume it directly without parsing an answer out of a sentence.
According to TypeSafe, three kinds: a yes/no answer with a probability attached, a choice among predefined options, and a score against a scale. A single selection can span up to 255 discrete options.
Because there is nothing to meter. Jev does not decode text autoregressively, so no output tokens are produced and none are billed. Only input is charged. This is a property of how the model works rather than a promotional discount.
TypeSafe reports end-to-end response latencies between 70 and 500 milliseconds. That figure is vendor-reported and measured on TypeSafe's own servers; it has not been independently reproduced.
On AI/ML API the model is served as typesafe/jev, which is Jev 1.13 at the time of writing. The "Jev Latest" name comes from the alias the model is listed under on other platforms, where it routes to the newest version in the Jev family — so what sits behind that alias can change as TypeSafe ships new versions.
TypeSafe states that schema matching is guaranteed and type errors are mathematically impossible, so Jev will always return a valid value from the set you allow. That constrains the shape of the answer, not its correctness — a perfectly well-formed decision can still be the wrong one.
No. The headline figures published at announcement — up to 193.6× faster and roughly 444× cheaper than conversational models — are vendor-reported, and the cost comparison was scored by TypeSafe against its own grading of other models' answers. No third-party reproduction has been published.
