The model cannot make up an answer: Jev selects from your list instead of writing text

Jev does not write text. It takes the state of your program and a question with a list of ready-made answers, then returns one option from that list and a probability for each option. It physically cannot return an answer outside the list.

That does not guarantee correctness. Jev can choose the wrong option, and a neatly formatted wrong answer remains wrong.

The company was founded in 2024 by Diogo Almeida, who worked on InstructGPT and ChatGPT at OpenAI. He now sells a model that does not write text at all.

Jev is not alone in guaranteeing form. OpenAI's Structured Outputs has maintained full JSON Schema compliance since August 2024, and fixed-list probability scoring has existed since 2020, when it was used to measure models on the MMLU test. What is new is different: previously, a separate classifier was fine-tuned for each task; now one model answers across multiple domains without training.

They sell probability, not a guarantee. Every Jev answer comes with a confidence score, and the promise is this: if it says 70%, the model is right in roughly seven out of ten cases. There is no public validation of that promise, and the company has not shown a single result on open benchmarks.

TypeSafe published its own run. Four practical tasks where the correct answer was defined as the averaged opinion of GPT-6 Astra and Fable 5.1: a measure of agreement with two other models, not correctness.

That is also where the showcase figures of 193.6x for speed and 444.6 for price come from. Competitors in that run had reasoning disabled, while the correctness reference had it enabled. In the round's press release, the same advantage is described as "up to 100 times," half as impressive as on the homepage.

**Outside observers measured something else.** Vercel calls Jev the fastest-adopted model in the history of its AI Gateway: within 18 hours, one tenth of paid teams had tried it; by the end of the first day, nearly 13%. That is twice the first-day figure for the GPT-5.6 family and six times that of Fable 5.1.

The waitlist can be bypassed. Since 16.09.2026, Jev has been available through Vercel AI Gateway with the model string typesafe-ai/jev. It can be installed directly with pip install typesafe-sdk, with a key obtained in the TypeSafe dashboard. Input costs $0.042 per million tokens, output is free, and a response arrives in 70–500 ms.

TypeSafe is more candid in its own documentation than on its storefront. It says Jev is poor at arithmetic, gets confused by hex and RGB color codes, will not judge whether two numbers are close, and reads dates as plain text. It also says the probability of a statement and the probability of its negation do not have to add up to one. For a model sold on calibrated probabilities, that is an inconvenient line.

Until someone outside the company measures Jev's probabilities, only the speed at which developers are picking the model apart speaks for it.