AI Tech Flash ── 2026-10-17
TypeSafe AI Releases Jev, an AI That Returns Decisions
A new API returns typed decisions with confidence, not text
What changes
According to TypeSafe AI, Jev does not write text. It returns typed decisions, meaning answers in a format fixed in advance.
Software can branch on the returned probabilities. The materials describe acting automatically when Jev's confidence is high and sending the case to a person when it is low.
Why it matters
TypeSafe AI argues that models tuned for chat tend to be overconfident and need humans in the loop. Jev attaches a confidence estimate to each answer, so software can decide when review is needed.
The price and speed figures are large. But according to Marktechpost, the main benchmarks were run by the vendor, and it advises testing on your own data.
What it is
Jev is the first public System One Model from TypeSafe AI. The company describes System One Models as a new class of AI model built for decisions inside software.
According to Marktechpost, Jev is transformer-based but is not a large language model, and it does not generate text. You send a state and typed questions, and it returns decisions with probabilities.
Who and where
Jev is made by TypeSafe AI, which says it was made in San Francisco. According to Marktechpost, it is available as a hosted API in early access behind a waitlist.
Weights, a parameter count, and a self-hosting option have not been published. OpenRouter's documentation lists the model ID typesafe/jev-1.14.
Why it was built
TypeSafe AI says Reinforcement Learning from Human Feedback (RLHF) tunes models to human preferences. It says this created chat, but also mode dropping, overconfidence, and a lack of reliability.
According to the company, these flaws mean large language models need humans in the loop. It says it built models meant for machines to use directly.
How it works
According to Marktechpost, one endpoint handles everything. You send a state and questions, and each question runs in parallel against the same state.
There are three question types. Choice picks one option from a list, Score rates against ordered levels, and Noul gives a probability from 0 to 1 that a statement is true. OpenRouter's documentation also lists Box, which locates something in an image.
Confidence is a value from 0 to 1, derived from the shape of the probability distribution. In the Marktechpost example, the top option scores 0.84 but confidence is 0.596. TypeSafe says training uses its own method, Reinforcement Learning for Calibrated Decisions (RLCD).
Key points
- Jev returns typed decisions with probabilities and confidence, not text.
- TypeSafe says input costs $42 per billion tokens and output is free.
- The speed and cost figures come from TypeSafe's own evaluations, not independent tests.
Conclusion
Jev is released as an AI specialized in decisions that machines use directly. TypeSafe AI claims it is 193.6x faster and 444.6x cheaper.
However, according to Marktechpost, those figures come from the company's own evaluations. Buyers need to test it on their own data before adopting it.
Quotes and sources
“System One Models are a new class of AI model built for decisions inside software.”── typesafe.ai, 2026
“Jev is transformer-based, but it is not a large language model. It does not generate text.”── MarkTechPost, 2026-09-19
“The main benchmarks are vendor-run. Test on your own data.”── MarkTechPost, 2026-09-19
Source articles
Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.