Jev
TypeSafe AI's System One model. You give it a state and typed questions; it returns decisions with probabilities, in one parallel pass, for $0.042 per million input tokens with output free.
Every agent harness makes dozens of small judgements per task. Which model should handle this? Is this tool call safe to run? Which queue does this message belong in? Did the agent actually do what it just claimed? Today those are usually answered by asking a frontier model in English and parsing the reply — slow, priced like prose, and free to invent an answer you never offered. Jev is built for exactly that layer and nothing else. You send a state and a set of typed questions; it returns every answer at once, each with a probability, and it cannot return a value outside the schema you gave it. It also cannot write your email, explain itself, or produce a single string — which is the trade.
Jev is not a smaller chatbot — it is a different shape of model: System One (fast, automatic judgement) to a frontier model's System Two (slow, deliberate reasoning), and the interesting builds use both, with Jev deciding and an LLM writing.
Guides
The three question types
| Type | What you give it | What comes back | Good for |
|---|---|---|---|
| Choice | Named options, each with a description | The selected option, plus a probability for every option | Routing, categorising, picking one value from candidates |
| Score | Ordered levels (e.g. low / medium / high) | A score across those levels with confidence | Severity, sentiment, buying intent, risk |
| Noul | A yes/no statement and instructions | The probability the statement is true | "Is this urgent?", "Was a refund requested?", "Is this claim supported?" |
You can ask several questions about one state in a single request. Adding questions barely changes the response time and costs only the extra input tokens, which is why a triage call can ask four things at once instead of four times.
At a Glance
| Factor | Detail |
|---|---|
| What it is | A System One model: typed, schema-constrained decisions with calibrated probabilities, sampled in parallel rather than token by token |
| Who makes it | TypeSafe AI. Jev is their first public model, in early access since 2026-09-15 |
| Price | $0.042 per million input tokens ($42 per billion); output free |
| Latency | 70–500ms end to end; measured evaluation times of 92–214ms in one tested run |
| Input context | 32K tokens |
| Limits | Choice cardinality up to 255; no image input yet; no text generation, no reasoning traces, no explanations |
| Access | console.typesafe.ai (early access) · OpenRouter typesafe/jev-1.13 (no waitlist) · Cloudflare Workers AI · LangChain and Pydantic AI integrations |
| Model id | jev-latest by default — an alias that resolved to Jev 1.13.0 when we checked. Pin it in production |
| Best for | The decision layer of an agent: routing, gating, triage, scoring, extraction from candidates, trace auditing |
| Not for | Chat, writing, code generation, anything that needs an explanation of why |
"Cannot hallucinate" is a claim about types, not about truth. Restricting the output stops the model inventing a category — it does not guarantee the category it picks is the right one. In AICodeKing's tests, a question about cafeteria hours correctly returned other when that option existed, and returned sales at 0.31 confidence when the options were only billing, technical and sales. So: always ship an other/unknown option and a human path, and treat calibration as something to measure on your own examples before you wire a decision to an action.
Tested breakdowns we've published
- Tested: routing, negation, a prompt-injection attempt and a browser run (AICodeKing, 2026-09-17)
- Live demos: gating a destructive tool call and scoring a sales lead (AiNow, 2026-09-19)
Related on This Site
- Cost calculator — Jev sits in its own vendor group, next to the frontier models you would otherwise use for these calls
- What is a System One model? · Jev (glossary) · Calibration
- Security center — gating tool calls, and why a classifier is a control rather than a guarantee
- Release entry — what shipped, with the vendor's own numbers
- Model & harness releases, Aug–Sep 2026 — where Jev sits among this month's launches
Sources, checked 2026-09-24: TypeSafe AI's System One announcement and quickstart, OpenRouter's Jev guide, and LangChain's harness write-up. Measured figures come from the tested video breakdowns we link, not from us. Jev is in early access and changing weekly.