TypeSafe AI released Jev last week, the first entry in what it calls the System One category: an AI decision model that answers structured questions about a piece of text or data rather than holding a conversation. Where a normal large language model takes a prompt and writes back prose, Jev takes a “state” object, such as a customer record or an article, plus one or more questions about it, and returns a probability or score for each. TypeSafe says the model was built by Diogo Almeida, an ex-OpenAI engineer who co-wrote ChatGPT’s core training techniques, and is meant for code to call directly rather than for people to read.
TypeSafe calls the category “System One” models, a reference to fast, intuitive human judgement. Simon Willison, who reviewed the API on the day of release, has argued that “decision models” is the more useful name, since what comes back is a typed, probabilistic decision rather than free text.
Inside TypeSafe’s AI decision model
TypeSafe trains Jev with what it calls Reinforcement Learning for Calibrated Decisions (RLCD), aimed at producing well-calibrated structured answers rather than fluent prose. Because each question is self-contained, with no conversation history and no memory retained between requests, a single API call can evaluate several questions against the same state in parallel, rather than generating tokens one after another the way a conventional LLM does.
TypeSafe’s own example illustrates the pitch: feed Jev a customer’s transaction history, account details and their most recent message, then ask whether they are requesting a refund. The answer comes back as yes or no with a confidence score, which the calling application can act on directly against whatever threshold it sets, without parsing free text.
Three question types, and what they cost
Jev accepts three kinds of questions, as developer Simon Willison detailed after testing the API directly: yes/no questions, which TypeSafe calls “Noul” questions after the Bernoulli distribution, return a confidence score between 0 and 1; choice questions return a probability distribution across a set of supplied options; and score questions return a value along a numeric range the caller defines. Willison also found that Jev charges only for input tokens, at $0.042 per million, cheaper than OpenAI’s GPT-5 Nano at $0.05 per million, with output left free.
Why the speed claim depends on which line you read
TypeSafe’s own comparison puts Jev at up to 445x cheaper than frontier models such as GPT-6 Astra, but the speed multiplier isn’t even consistent within the coverage: one outlet’s headline states Jev is up to 193 times faster, while the same article’s body text puts the figure at 194 times faster. Neither number comes from an independent benchmark; both trace back to TypeSafe’s own comparison chart. Readers who want to check the mechanics rather than the marketing can go straight to TypeSafe’s documentation, including its citation-checking cookbook and a public example on GitHub, jev-leftpad, that runs the API against a minimal test case.
What to watch next
No independent lab has published a benchmark of Jev against GPT-6 Astra or any other frontier model, and nothing in the coverage so far cites one. The real test of the speed and cost claims will come once developers running examples like jev-leftpad, or building their own intent-routing pipelines against Jev’s API, publish comparisons that don’t originate from TypeSafe’s own documentation.
