Jev

What it costs

• API input$42 per billion tokens$0.042 per 1M tokens
• API outputTypeSafe's blog: "too cheap to meter"Free

Official pricing ↗ · checked September 28, 2026

What it is

Jev is made by TypeSafe AI, a San Francisco lab founded in 2024 by Diogo Almeida, Erik Gafni and Sasha Sheng. Almeida previously worked at OpenAI on RLHF, InstructGPT, ChatGPT and GPT-4.

TypeSafe calls Jev a "System One model": instead of generating text it returns a typed decision with calibrated probabilities (plus a confidence score for choices and scores), meant for software to act on. It cannot chat, write emails, create images or generate code.

Developers call it through a REST API with Python and JavaScript SDKs, and it is also in beta on OpenRouter. A request (64k-token budget) asks Jev to choose between up to 255 options, score something on a scale, or give a yes/no probability.

Benchmarks and evidence

• Maker’s claim TypeSafe: 40–200x faster than frontier LLMs on "System One tasks", and up to 444.6x cheaper on four internal workflows it built itself ↗
• Maker’s claim TypeSafe says it will not publish a standard benchmark table ↗
• Independent Check Point: prompt injection flipped Jev's verdicts in all 9 attack settings it tried ↗
• Independent, informal PriorBench: a pre-registered test by an anonymous researcher, with raw data public ↗

Limitations

• "Cannot hallucinate" only means the answer is always one of your options; a well-formed answer can still be the wrong one, and accuracy is the open question (The Register).
• You have to define the options in advance, and decide yourself what a given confidence score means for your use case (TechCrunch).
• No paper, architecture details or weights have been published (Wikipedia).
• Security researchers at Check Point showed its verdicts can be manipulated with prompt injection.