Twenty questions developers actually ask about the model, answered from launch documentation and early-access reports.
As of: September 18, 2026. Answers reflect launch documentation from September 15, 2026 and early-access reports. Where
something has not been stated publicly, that is said rather than guessed at.
Everything people ask
What is Jev AI?
Jev AI is the first model from TypeSafe AI, launched September 15, 2026. Rather than generating text, it returns a typed decision — a choice from up to 255 options, a score on a 2–10 scale, or a yes/no probability — in 70 to 500 milliseconds, with calibrated confidence attached.
Who made Jev AI?
TypeSafe AI, founded by Diogo Almeida, who co-invented ChatGPT while at OpenAI. The company spent roughly two years in stealth before launching Jev AI and its RLCD training method.
Can Jev AI hallucinate?
It cannot return a value outside the schema you define, so fabrication in the LLM sense is structurally impossible. It can still return the wrong valid answer — a mistake within your option list, not an invention outside it.
Can Jev AI write code or prose?
No. It was never trained to generate text at all. If you need language output, you need a language model — the two tools do not overlap.
How much does Jev AI cost?
$0.042 per million input tokens, which is $42 per billion. Output tokens are free. On TypeSafe's benchmark that works out to roughly $0.0004 per decision, against $0.0304 and $0.0836 for frontier alternatives.
How fast is Jev AI?
It answers in 70 to 500 milliseconds. The speed comes from a parallel sampler that produces all outputs in a single query rather than emitting tokens sequentially.
How do I get Jev AI access?
Early access is waitlisted at typesafe.ai. Developers are invited off the list over time and notified by email. No public free-credit or trial programme was described at launch.
What are Choice, Score and Noul?
The three Jev AI question types. Choice returns one option from up to 255. Score returns a position on a 2 to 10 level scale. Noul returns a yes/no probability. Every Jev AI response carries calibrated confidence.
Is Jev AI an LLM?
No. It is a neural network, but it is not a large language model — it has no text generation capability whatsoever. The output is a typed decision, and that is architectural rather than a restriction layered on top.
Why does it not generate text?
Because generating text sequentially is what makes LLMs slow and expensive. Removing that requirement is what allows parallel sampling, sub-second latency, and pricing measured per billion tokens.
Could I just use structured output on an LLM?
You can, and many teams do. The difference is cost and latency: LLM structured output still generates tokens sequentially and is still billed per million, so the economics of running it on every item in a pipeline are entirely different.
Why price per billion tokens?
Because the numbers would be unreadable otherwise. At $0.042 per million, quoting per billion keeps the figure above one dollar and makes the comparison against per-million LLM pricing obvious.
Are output tokens really free?
Yes. The founder's description is that they are too cheap to meter, which follows from the output being a typed value rather than a long generated sequence.
Is the pricing subsidised?
TypeSafe is candid that it cannot prove otherwise, and says long-term sustainability will have to be demonstrated. It expects prices to fall rather than rise.
Is there a free trial?
No free-credit or trial programme was described in the launch materials. Access is through the waitlist, and pricing applies once you are in.
How long is the waitlist?
TypeSafe has not published a timeline. Early-access reports from the first days suggest developers are pulled off the list in batches rather than all at once.
Do I write prompts?
Not in the LLM sense. You declare a schema — the options, the scale, or the binary — and supply context. The model fills in a value inside that schema.
What does the API return?
A typed value inside your declared schema, plus a calibrated confidence figure. Never free text, never tokens you have to parse.
How many options can a Choice have?
Up to 255. Beyond that you need to restructure the question, typically by splitting it into a coarse routing step and a finer follow-up.
What scale does Score use?
A 2 to 10 level scale that you define. It returns a position on that scale rather than an arbitrary number, which keeps outputs comparable across calls.
The short version on Jev AI
If you only take one thing away: this is not a smaller, cheaper chat model. It is a different kind of
component. It answers questions whose answers are values — which of these, how much, yes or no — and
it does so fast enough to sit in a request path and cheaply enough to run on everything rather than on
a sample.
The decision about whether to adopt it is therefore not a model comparison. It is an audit of your own
pipeline: how many of the calls you currently make to a language model are actually classification,
scoring, or gating dressed up as a conversation? Whatever that number is, it is the size of the
opportunity, and it is usually larger than people expect.
The rest is ordinary engineering. Define the schema carefully, calibrate on data you already have
labels for, set thresholds from what you observe rather than what you hope, and keep a person on the
uncertain band. None of that is novel — it is just that the economics finally allow it at a scale
where it changes what is worth automating.