Jev by TypeSafe AI

This page covers Jev by TypeSafe AI — not the rapper or the Japanese encephalitis vaccine.

What Jev is

Jev is a decision model from TypeSafe AI, announced on September 15, 2026.^[1] Where a large language model generates text token by token, Jev takes a piece of state plus a typed question and returns a typed answer — one of a set of options, a score on an ordered scale, or a yes/no probability — in a single pass.^[2]

TypeSafe AI

TypeSafe AI is the company behind Jev, led by founder Diogo Almeida, who previously worked at OpenAI on the methods behind ChatGPT.^{[1] [4]} The company describes Jev as its flagship model and the first of a class it calls System One models.^[2]

System One

System One models are a category of AI models built to return fast, structured decisions that software can use directly, rather than generated prose. The name references the fast, intuitive “System 1” thinking described in Daniel Kahneman's Thinking, Fast and Slow.^[2]

How it works

Jev's interface is three primitives, which can be mixed in one request against the same state:^[2]

• Choice — picks one option from a set you define and returns the probability of each option.
• Score — places the input on an ordered scale you define and returns a probability-weighted position.
• Noul — a yes/no answer returned as a probability from 0 to 1.

Jev — confirmed limits

TypeSafe's model documentation confirms the following operating limits and billing details.^[3]

Model version

Jev 1.13^[3]

Context window

64,000 tokens — 32,000 reserved for state, plus the longest single question^[3]

Throughput

Up to 100K tokens/s and 40 requests/s^[3]

Input

Text only — strings, JSON, or arrays of text. No image, audio, or video input.^[3]

Language

English is the primary training language. CJK is supported but does not perform equivalently.^[3]

Billing

$42 per billion input tokens ($0.042 per million); output tokens are free.^[3]

Schema guarantees

Because the answer must be one of the values you defined up front, Jev structurally cannot return a malformed type or an out-of-set option. That is a format guarantee, not a correctness guarantee — the model can still pick the wrong option from your set.^[2]

Calibrated probabilities

TypeSafe trains Jev with RLCD (Reinforcement Learning for Calibrated Decisions) so that its reported probabilities reflect real uncertainty in aggregate — a 0.8 means it is right about 80% of the time across similar cases. Calibration is measured across groups of predictions and does not guarantee any single answer is correct.^[1][2]

How it differs from a general-purpose LLM

An LLM returns tokens — prose, code, or JSON — that a caller then has to parse and validate. Jev returns a typed value plus a probability distribution you can branch on directly. It accepts text input only (strings, JSON objects, and arrays of text); it does not generate replies or explanations.^[2][3]

TypeSafe reports end-to-end latency of 70–500 ms (self-reported)^[5] and input pricing of $0.042 per million tokens with output free.^[3] Jev is a hosted, closed-weight model; it is available through TypeSafe's API and through gateways such as OpenRouter.^[4]

How it compares with the Decisions API

OpenAI's Decisions API and Jev target the same job: return a decision instead of text. The main difference is specialization — Jev is a purpose-built, non-generative decision model, while the Decisions API folds decision output into a general model platform. See the Decisions API vs Jev comparison, or read about OpenAI's Decisions API and the broader decision API comparison.