Jev AI is a model from TypeSafe AI that returns typed, probabilistic decisions instead of generated text. A "typed decision" means an answer with a shape you define in advance, such as a choice from a list, a number, or a yes/no, each with a confidence score. TypeSafe released Jev in limited early access on 15 September 2026.
Jev is not a chatbot, and it does not write sentences. You give it context and a specific question, and it hands back a decision your code can act on directly. Think of it as a fast decision-maker that sits inside your software, not a writer you read.
How Jev Is Different From a Large Language Model (LLM)
A large language model generates text one token at a time, with each token predicted from the last. Jev instead produces its answer in a single pass, without generating text word by word. This single-pass, non-autoregressive design is the core difference, and it is why Jev can return a decision far faster than a chat model.
That speed is the headline in most launch coverage. According to one launch analysis, "Typesafe claims this makes Jev up to 100 times faster and 100 times cheaper than conventional LLMs for certain tasks, while remaining competitive on intelligence." These are the company's own figures, so treat them as vendor claims rather than independently verified results.
What "System One" Means
TypeSafe calls Jev a "System One model." The name borrows Daniel Kahneman's split between System 1 thinking, which is fast and intuitive, and System 2 thinking, which is slow and deliberate.
In software terms, the two map to different jobs. You use an LLM for open-ended reasoning and writing, and a model like Jev for fast, bounded decisions along the way.
How Jev Works: State and Typed Questions
You send Jev two things. The first is "state" — the context it should judge, such as a support ticket, a message, or a JSON object. The second is one or more "typed questions," each with a defined set of allowed answers.
Jev then returns values your code can branch on, such as a category, a score, or a true/false. A decision is only as good as the context behind it, so the quality and structure of that state matter. Much of that state comes from the live web, and Olostep's Answers endpoint is one way to pull structured JSON from the web to feed a decision step.
The Three Kinds Of Typed Answers
Jev answers within an answer space you define, and it cannot return a value outside it. Each answer comes with probability or confidence signals you can threshold on.
| Question type | What it returns | Example |
|---|---|---|
| Choice: pick from a fixed list | One option from your set | Route a ticket to billing, technical, or sales |
| Score: a number in a range | A numeric value | Rate urgency from 1 to 5 |
| Yes/No: a boolean | true or false | Does this action need human approval? |
How Jev Is Trained (RLCD)
TypeSafe says it trains Jev with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD), which aims to make the model's confidence scores well-calibrated. This is a company-defined term, and there is no independent academic validation of the method yet. Read it as TypeSafe's description of its own approach.
Who Built Jev, And Why It Got Attention
Jev was built by TypeSafe AI, a San Francisco company founded in 2024. Its founders are Diogo Almeida (CEO), Sasha Sheng, and Erik Gafni.
Almeida spent about four years at OpenAI, where he helped develop reinforcement learning from human feedback (RLHF) and co-authored InstructGPT, the research behind ChatGPT and GPT-4. That background is a big reason the launch drew attention. Per Jev's public launch details, "It was released in limited early access on 15 September 2026, alongside the announcement of a US$40 million seed round led by DCVC."
The name is a nod to the 19th-century economist William Stanley Jevons. His Jevons paradox describes how making a resource cheaper can increase how much of it people use, and TypeSafe expects cheaper machine intelligence to follow the same pattern.
What Jev Is Good For: Classification, Routing, Scoring, And Safety Checks
Jev fits narrow decisions that software needs to make over and over. Common jobs include classifying a support ticket, routing a request to the right team or model, scoring how urgent a task is, and gating a risky tool call with a safety check.
These are the steps where teams often prompt an LLM and then parse its JSON reply. Jev is built to make that specific decision directly, which removes the parse-and-validate step. For example, in a multi-agent research pipeline, a fast decision could pick which source to trust or which agent should run next.
Where Jev Fits In An AI Agent Loop
Inside an AI agent, Jev handles the fast decision points while an LLM handles generation. The pattern is simple: the agent gathers context, makes a decision, and then acts.
A model-routing step can send easy requests to a small model and hard ones to a larger model. Those same decision points show up in scheduled web research agents, where each run has to decide what to fetch, what to keep, and when to stop.
Speed And Cost: What TypeSafe Claims
Jev's pricing is a concrete, published fact, while its headline speed and cost multiples are vendor-reported. According to TypeSafe's launch announcement, "Input tokens: $0.042 / MTok ($42 per billion tokens). Output tokens: FREE (too cheap to meter)."
The performance ratios need more care. As LangChain's engineering write-up notes, "The company reports up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks." TypeSafe itself says these figures are at the high end of real-world gains, so treat them as claims to test on your own workload, not guarantees.
Limitations: A Typed Answer Is Not Always A Correct Answer
Type safety guarantees the shape of the answer, not its truth. Jev cannot return a value outside your defined answer space, but it can still misread the context and hand back a valid-but-wrong result. As Vercel's developer guide puts it, "An answer can fit the allowed values and still misinterpret the evidence."
A few other limits are worth knowing before you rely on it:
- Text-based inputs: Jev's current inputs are text, so you prepare a text description or transcript before asking about media.
- Not a writer: it does not generate explanations or code, so you still need a generative model for those.
- Test before you trust: check its decisions against known outcomes before you let them trigger real actions.
Jev And The Rest Of The AI Agent Stack
Jev is one piece of a larger stack, not a whole system. An agent usually has to retrieve fresh data, decide, and then act, and Jev covers the decide step.
A fast decision is only as good as the state it receives, and much of that state comes from the live web. Olostep positions itself as infrastructure for AI agents, turning pages into clean, structured context. Because models read cleaner inputs better, delivering that context in AI-ready data formats like Markdown and JSON helps a decision step work well. In this view, data infrastructure and decision models are complementary, not competing.
Frequently Asked Questions
Is Jev An LLM?
No. Jev is a decision model that returns typed values with confidence scores rather than generated text, so it is used differently from a chat LLM.
Does "Can't Hallucinate" Mean It's Always Right?
No. Jev cannot produce an answer outside your defined schema, but it can still misjudge the evidence and return a value that is valid in shape yet wrong.
Is Jev Free?
Jev launched in limited early access, and TypeSafe lists input tokens at $0.042 per million with output tokens free. Prices can change, so check current rates before you build on them.
Will Jev Replace LLMs?
No. Jev is complementary: teams use an LLM for open-ended reasoning and generation, and a decision model like Jev for fast, bounded decisions.
Can Jev Read Images Or Other Languages?
Jev's current inputs are text-based, so you prepare a text description or transcript first. It is not designed to take images or audio directly.