explainer · updated 2026-09-21

# What is Jev? The System One model, explained

**Jev** is an AI model from TypeSafe AI that decides instead of writing. Give it any text or JSON and a few typed questions: it answers all of them at once, with a probability for every possible answer. It is the first *System One model*.

- **maker**: TypeSafe AI
- **launched**: 2026-09-15
- **answers**: choice · score · noul
- **price**: $0.042 / M input tokens
- **median latency**: 0.44–0.48 s
- **model**: jev-1.13.0

## What is a System One model?

An AI model that answers typed questions with probabilities, all in one pass, and never writes text.

System 1 · fast

### One look, one answer

**Jev**: every question answered at once, as probabilities.

System 2 · slow

### Step by step, in words

**LLMs**: text one token at a time, reasoning by writing.

The name borrows the fast, intuitive System 1 of Daniel Kahneman's [*Thinking, Fast and Slow*](https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow). Most decisions software makes are that kind: spam or not, which team, safe to run.

### typed

Answers only with your options. Nothing to parse.

### parallel

Every question in one pass. Ten take as long as one.

### probabilistic

Odds for every answer, not just a pick.

### cheap

No text is generated, so output is free.

## How does Jev work?

State in, decisions out. One API request carries the input and every question, and each question is answered on its own.

*The ticket above, taken apart. Each question runs in its own lane.*

- **state**: What the decision is about: any text or JSON. The context window is about 32,000 tokens, text only.
- **question**: A typed ask, `choice`, `score` or `noul`, with an instruction.
- **criteria**: The possible answers, described in plain words.
- **independence**: Questions never see each other's answers. Combine them in code.

## What are choice, score and noul?

Jev's three question types: pick one option, place on an ordered scale, or answer yes or no.

### `choice` → match

Pick one of up to 255 options.

request state + question

- **message**: Can I move my table from Friday to Saturday night?

What does the customer want?

choice — book — change — cancel — ask — with descriptions

- **for**: routing, intent, next action
- **fields**: `choice`, `probabilities`
- **note**: Probabilities sum to 1, so add an other option.

### `score` → sort

Place it on 2 to 10 ordered levels.

request state + question

- **review**: Food was fine and the service slow, but the view made up for it.

How positive is the review?

score — Negative — Mixed — Positive — Glowing

- **for**: severity, quality, ranking
- **fields**: `score`, `probabilities`
- **note**: The score is the expected level: sort by it.

### `noul` → if

Yes or no, as the probability of yes.

request state + question

- **email**: Your parcel is held at customs. Pay the $2.99 release fee within 24 hours: parcel-release.help/pay

Is this a phishing attempt?

noul — yes — no

- **for**: flags, checks, gates
- **fields**: `noul`
- **note**: Short for Bernoulli; boolean in the Vercel AI SDK.

## How is Jev different from ChatGPT and other LLMs?

An LLM writes its answer token by token, and your code parses it. Jev is not an LLM: it writes nothing and scores your answers, all at once.

LLM · one token at a time

Jev · every answer at once

|  | LLMs | Jev |
| --- | --- | --- |
| Output | Text to parse | Probabilities over your answers |
| Writes text or code | Yes | No |
| Reasons step by step | Yes | No: one look per question |
| Latency (median) | 0.53–2.21 s | 0.44–0.48 s |
| Cost per 1,000 decisions | $0.023–$1.5 | $0.029–$0.043 |
| Possible answers | Anything, including made-up ones | Only the options you defined |
| Probabilities | Only if asked to write them | Built in, trained to be calibrated |
| Best at | Open-ended work | Many fast decisions over known answers |

Measured on the Jevals boards: Jev and 6 LLMs, release 2026-09-18.

### Is Jev just a classifier?

Close: it does zero-shot classification, and critics say so. The difference from a trained classifier: you describe the options in words and change them on any request.

|  | Trained classifier | LLM with structured outputs | Jev |
| --- | --- | --- | --- |
| Training data | Needed | None | None |
| New answers | Retrain | Edit the prompt | Edit the request |
| Output | Scores per label | Text that should parse | Probabilities per answer |
| Speed and cost | Fastest | Slowest | Fast |

## What do Jev's probabilities mean?

They are meant to be calibrated: answers given at 80% should be right 80% of the time. That lets code act alone when Jev is sure, and ask for help when it isn't.

sure: act

unsure: ask

*Same pick, different confidence. Illustration.*

yes/no

gap **5.0** pts

pick-one

gap **9.8** pts

rubric scores

gap **19.7** pts

*Jev's real answers, release 2026-09-18: stated confidence across, accuracy up, perfect on the diagonal. Asked for probabilities, the best LLM on each board comes closer (Gemini 3.8 Flash 2.0, DeepSeek V4.1 Flash 2.8, GLM-5.3 12.9).*

## What is Jev used for?

Decisions with known answers that software makes often: route, check, rank, choose.

- **1 · propose** Code or an LLM lists the options.
- **2 · decide** Jev scores them against the state.
- **3 · gate** Code checks the confidence.
- **4 · act** Sure: do it. Unsure: escalate.

choice

### Routing

Send each ticket, email or prompt to the right place.

noul

### Guardrails

Approve an agent's command, or ask a human.

choice · noul

### Moderation

Publish, redact or reject.

score

### Judges and linters

Grade outputs against a rubric, in place of an LLM-as-a-judge.

score

### Ranking and reranking

Score every search result, then sort.

choice

### Classification

Label a million rows by topic, intent or sentiment.

choice

### Agents

Pick the next action or the element to click.

choice

### Real-time loops

Game bots deciding several times a second.

choice

### Extraction

Code finds candidates, Jev picks one or none.

### Limitations: when not to use it

- Writing anything: replies, summaries, code.
- Answers you can't list in advance.
- Arithmetic or multi-step reasoning.
- Security boundaries: text in the state can sway it.
- Its [documented weak spots](https://docs.typesafe.ai/model-jaggedness/jev-1.13): counting, dates, negation.

## How does Jev score on benchmarks?

Tied for first on yes/no and tied for second on pick-one, for a fraction of most LLMs' cost. On rubric scores, no model clearly beats guessing.

| Board | Test | Jev Decision Score | Jev accuracy | Best other model |
| --- | --- | --- | --- | --- |
| [`noul`](https://jevals.com/noul/) | PubMedQA | **69.0** | 91.3% | Gemini 3.8 Flash 73.0 |
| [`choice`](https://jevals.com/choice/) | Banking77 | **67.8** | 79.7% | Gemini 3.8 Flash 74.1 |
| [`score`](https://jevals.com/score/) | HelpSteer2 helpfulness | **9.2** | 41.3% | GLM-5.3 7.8 |

Scored against human labels, five runs per question; orange marks each board's top 2. Head to head: [Jev vs Gemini 3.8 Flash](https://jevals.com/compare/jev-vs-gemini-3.8-flash/).

## Who makes Jev?

TypeSafe AI, a San Francisco startup led by CEO Diogo Almeida, a former OpenAI researcher on InstructGPT.

- 2024 TypeSafe AI founded, in stealth
- 2026-07 Talks on what comes after RLHF
- 2026-09-15 Jev launches
- 2026-09-16 On Vercel AI Gateway
- 2026-09-18 [First independent benchmark](https://jevals.com/notes/2026-09-18/)
- **training**: RLCD, Reinforcement Learning for Calibrated Decisions. Unpublished, like the architecture.
- **name**: After the [Jevons paradox](https://en.wikipedia.org/wiki/Jevons_paradox): make something cheaper and use grows.
- **access**: TypeSafe's API (waitlist), or Vercel AI Gateway as `typesafe-ai/jev`.
- **funding**: $40M seed round led by DCVC.

## What is a Jev-type model?

Any model that answers choice, score and noul with probabilities. Jev itself is closed; adapters and open-source alternatives now copy its interface.

| Model | What it is | Open | On Jevals |
| --- | --- | --- | --- |
| [Jev](https://jevals.com/models/jev/) | The native System One model | No | Listed |
| LLMs via an adapter | An LLM prompted for the same probabilities (system-one-adapter, AI SDK `evaluate`) | Varies | [6 listed](https://jevals.com/models/) |
| openjev | Option probabilities read from an open Qwen model | MIT | Coming |
| decider-2b | A 2B Qwen model fine-tuned for the three types | Apache-2.0 | Coming |
| DiffusionGemma | Google's diffusion model filling an answer template (open-jev) | Open weights | Not yet |
| GLiClass, GLiNER | Small zero-shot encoder classifiers, the prior art | Apache-2.0 | GLiClass coming |

## What are the criticisms of Jev?

Mostly fair: it is close to a zero-shot classifier, and most claims about it are the vendor's own.

- **"Just a zero-shot classifier."** Partly true: encoders like GLiClass came first. New is one hosted model, three types, calibration as a goal.
- **"Can't hallucinate."** It can't invent an answer, but it can pick the wrong one, confidently.
- **Vendor benchmarks.** TypeSafe's [Workflow Evals](https://evals.typesafe.ai/) grade agreement with two LLMs, not human labels.
- **A black box.** No paper, no parameter count, no published calibration.
- **The price.** Critics suspect a subsidy; TypeSafe says it is profitable.
- [boards**Jev vs 6 LLMs**](https://jevals.com/noul/)
- [explorer**Every answer, every model**](https://jevals.com/explore/)
- [methodology**How Jevals scores**](https://jevals.com/methodology/)

Facts checked 2026-09-21 against TypeSafe's and Vercel's public pages. Jevals is independent of TypeSafe AI.

Source: https://jevals.com/what-is-jev/