Overview
Jev is TypeSafe AI's first System One model, announced on its blog on 15 September 2026 as "available today in early access". It does not generate prose. A request carries a state plus one or more typed questions, and the response is the answer itself: the documentation describes it as "Send state and typed questions; get structured answers your code can use directly. No text generation, no parsing."
Three primitives define what a question can ask for: Choice returns a choice with probabilities and a confidence, Score returns a score the same way, and Noul returns a value between 0 and 1. TypeSafe says the probabilities are calibrated — "optimized against outcomes to reflect uncertainty" — while being explicit that calibration "does not guarantee that an individual answer is correct".
The trade the company is selling is latency and price rather than breadth. Its launch post puts end-to-end response time at 70 ms to 500 ms against 3 to 329 seconds for frontier models on the same System One tasks, which it summarises as "40x-200x faster", and it prices input at $0.042 per million tokens with output tokens free — "too cheap to meter". The post compares Jev against GPT-5.6 Terra, GPT-6 Astra, Fable 5.1 and DeepSeek models on its workflow evaluations but publishes no numeric benchmark table, so this page carries none.
It is a narrow model by design. Input is text only — "String, JSON object, or array of text values" — with images, audio and video unsupported, English the best-served language, and TypeSafe's own docs listing nine failure modes including mathematical reasoning, date comparisons and literal interpretation of instructions. The intended home is the decision points inside software: routing, classification, triage and scoring, where a fast predictable typed answer matters more than an explanation.
| Released | 2026-09-15 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Context | 64K |
| Max output | A typed value with probabilities and a confidence — no generated text |
| Modalities | Text |
| Status | Early access |
Pricing
| Input | $0.042 / 1M tokens |
|---|---|
| Output | $0.00 / 1M tokens |
Charged per input token; output tokens are free. TypeSafe also quotes the input price as $42 per billion tokens.
Strengths
- Returns a typed value with calibrated probabilities, so there is no output to parse or schema-validate
- TypeSafe reports 70 ms–500 ms end-to-end response time, which it frames as 40x–200x faster than frontier LLMs on these tasks
- $0.042 per 1M input tokens with output tokens free, charged on input only
- Documented rate limits of 250,000 tokens/second and 1,200 requests/minute
- Stable aliases (jev-latest, jev-preview) resolve to the pinned jev-1.13.0
Best for
- Routing a request to the right branch, tool or model inside an application
- Classification, triage and scoring steps in an agent or automation pipeline
- Replacing an LLM call whose only job was to emit a single structured field
- High-volume decision points where per-call latency and cost dominate
How to access
| Provider | Model ID |
|---|---|
| TypeSafe AI ↗ | jev-1.13.0 |
FAQ
What is a System One model?
TypeSafe's term for a model "built to make fast, structured decisions that software can use directly". Where an LLM generates natural language, a System One model evaluates typed questions against a state and returns constrained typed answers with probabilities.
Which model id do I call?
jev-1.13.0. TypeSafe also publishes two aliases — jev-latest (the stable default) and jev-preview — both of which currently resolve to jev-1.13.0.
How big is the context window?
64K tokens. TypeSafe's docs explain the split: the 64K budget covers the state plus all questions combined, while a 32K limit applies to the state plus the longest single question.
What does it cost?
$0.042 per million input tokens, quoted by TypeSafe as $42 per billion tokens. Output tokens are free — billing is on input only.
What can't Jev do?
It does not generate text, and it takes text input only — no images, audio or video. TypeSafe's own "jaggedness" page lists nine failure modes, among them mathematical reasoning, date comparisons and taking instructions literally, and notes that English is handled best.