TypeSafe AI's Jev answers with probabilities and bills nothing for output
reported by 2 outlets· 2 articles · confidence: medium · first seen 2026-09-19 18:41 UTC
What this means for you
If you pay a text model to label records at volume, this is worth testing: input costs $0.042 per million tokens, and there is no output charge because nothing is generated. The catch is whether the confidence scores hold up — no independent evaluation exists yet. Pilot on your own data before rerouting a pipeline.
TypeSafe AI has released Jev, a model that takes text in and returns numbers instead of prose: the probability a statement is true, a distribution across options you supply, or a score on a range you supply. The vendor calls the category a "System One" model. It is priced at $0.042 per million input tokens with output free, undercutting OpenAI's GPT-5 Nano at $0.05 per million. One call takes a record plus as many questions as fit the context. Benchmark figures are the vendor's own, with no independent evaluation or dates given.
Models in this story
Key facts
- ·Jev returns floating-point numbers — probabilities for yes/no questions, a distribution across supplied choices, and scores along a supplied range — rather than generated text. source
- ·Input is priced at $0.042 per million tokens and output tokens are free. source
- ·OpenAI's GPT-5 Nano charges $0.05 per million input tokens, making Jev's input cheaper. source
- ·The API accepts a "state" object — a string, an array of strings, or name-value pairs — with multiple questions in a single call. source
- ·TypeSafe calls the yes/no question type "Noul", which its CEO confirmed on Hacker News is short for Bernoulli. source
- ·The release ships with vendor benchmarks carrying stated caveats and documented limits; no independent evaluation is cited. source
What the sources say
- Simon Willison — Walks through the API shape, the three question types and the pricing comparison with OpenAI's cheapest model.
- MarkTechPost — Brief announcement post; the substance sits behind the release's own benchmark and limits documentation.
Sources
The original reporting. Follow these — they did the work.
- MarkTechPostTypeSafe AI Releases Jev: A System One Model That Returns Typed, Calibrated Decisions Instead of Text2026-09-19
- Simon WillisonJev introduces a new shape of LLM - System One, aka Decision Models2026-09-21