TypeSafe ships a model that answers only with a decision and a confidence number

The step this is for is the one in your agent that ends in a yes or a no: retry or fail, allow or block. Jev answers that kind of question and nothing else, because it cannot produce strings, so it cannot chat, draft, explain or write code ^[4]. Answers come back as a choice, a score or a yes-no probability, each with a calibrated confidence number attached, and TypeSafe says the confidence number is the part that matters ^[2].

Diogo Almeida worked on the instruction-following research behind ChatGPT at OpenAI, and the four years since left him disappointed by what it became ^[9]. "The problem is we are optimizing for human language," he told TechCrunch ^[8]. TypeSafe trains Jev entirely on synthetic data it generates itself, using a method it calls reinforcement learning for calibrated decisions ^[19].

The company's argument for the confidence number is narrow and testable: a model might do a job 95% of the time, and unless it also says when it is in the other 5%, nothing can be automated around it ^[3]. The schema fixes the option set before the call goes out, so the model cannot invent a fifth option when asked to pick among four, which TypeSafe calls a structural property and not a measured result ^[18].

Calibration quality is also the part no outside party has measured. TypeSafe's own model capabilities team built the workflow evaluations ^[13]. The comparison against language models uses an average of GPT-6 Astra and Fable 5.1 as the reference answer, which the company concedes biases the results towards OpenAI and Anthropic ^[14]. Its figures for rival models come from OpenRouter traffic, which may send more complex queries to better models ^[15]. The company has chosen not to publish results against public benchmarks, arguing that users should build their own ^[16].

The day-one adoption number needs the same reading. Jev's near-13% share of Vercel's paid teams was twice that of the GPT-5.6 family and more than six times Fable 5.1's ^[11], which works back to roughly 6.5% for the GPT-5.6 family and under about 2.2% for Fable 5.1 ^[21][22]; every other recent launch stayed below 7% after a full day ^[12]. A gateway share counts teams that changed a config value. At $0.042 per million input tokens and nothing charged for output ^[6], that experiment is close to free, and the figure covers the first 24 hours ^[10].

A team with a fixed option set and a ground-truth log can check Jev's confidence numbers against its own traffic inside a week and automate the cases above its threshold. Without the log, you are buying latency and price, and should price it that way: 500 milliseconds at Jev's slow end against three seconds at the fast end of TypeSafe's range for frontier models is about a sixfold gap ^[7][23]. A team whose step ends in a sentence needs a language model regardless.

TypeSafe named the model after William Stanley Jevons, the nineteenth-century economist whose paradox holds that a falling price for a commodity tends to raise consumption ^[17].