TypeSafe AI's Jev outputs numbers, not text, at $0.042/M input

TypeSafe AI launched a language model that does not write. Jev takes text in and returns floating point numbers, and Simon Willison, writing on his blog, calls the new category 'decision models' (TypeSafe's own label is 'System One models').

The pricing is the eye-catch. Jev charges $0.042 per million input tokens; output is free, because there is no text output. That undercuts OpenAI's GPT-5 Nano at $0.05 per million, and Willison's post calls it 'very fast, and really cheap.'

The model answers three shapes of question. Yes/no queries return a confidence between 0 and 1. Choice questions return a probability distribution across the supplied options. Score questions return a numeric rating. 'Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out,' TypeSafe writes in its launch material.

TypeSafe AI itself came out of stealth a week earlier, on September 15, with a $40M seed round led by DCVC. The founders are Diogo Almeida, formerly of OpenAI, along with Erik Gafni and Sasha Sheng, according to Truefoundry's write-up.

Willison's post is not uncomplicated praise. He calls the shift 'a regression even further towards black box machine learning systems': 'put in all the text you want, the only thing you're going to get back is a floating point number.' When he ranked cities on 'Good city?', Cupertino came out on top and East Palo Alto at the bottom, the kind of ordering he flags as hard to interrogate in a system that only returns a score.

He shipped an llm-typesafe plugin for his LLM CLI the next day. Two of the researchers we follow circulated the post as it landed. Community projects followed within days: jevchat, jev-leftpad, jev-2048, and an open-weight riff called Kev built on Qwen.