What Jev is: a System One decision model that doesn't generate text
The ChatGPT co-creator's new model Jev answers typed Choice/Score questions with calibrated probabilities at $0.042 per million tokens and 70-500ms latency, aimed at moving 90% of classification calls from System 2 pricing to System 1.
200x faster. 400x cheaper. And it can't hallucinate - because the answer space is defined before it even runs The chatgpt co-creator left OpenAI, spent 2 years in silence, and built a model that doesn't generate text at all Jev is a "System One" model. your code asks it typed questions - Choice, Score, Noul - and it returns calibrated probabilities for every option in one parallel pass. no tokens. no parsing. no retry loops $0.042 per million input tokens. output included free. 70 to 500ms latency The name references Jevons paradox: when intelligence gets this cheap, you don't optimize the same use cases. you automate an entire class of decisions that were never worth the API call before Normal LLMs generate your answer token by token, then you pray the string is valid. Jev generates probabilities for each option in parallel - the answer space is defined before it even runs 90% of production LLM calls are classification. we've been paying System 2 prices for System 1 work this whole time
This post quotes another post: x.com/0xaigoat/status/2102027640752177195