Jev: TypeSafe's Model That Writes No Text and Returns Typed Decisions in 70 Milliseconds
TypeSafe has released Jev, a model that writes no sentences: you hand it your app's state and it returns typed decisions with probabilities, at $0.042 per million input tokens. Within 24 hours it became the fastest-adopted launch in Vercel's AI Gateway.
What Jev Is and Why It Is Not Just Another LLM
A language model takes text and returns text. Jev does something else: it takes your application's state in plain language plus typed questions, and returns a structured decision — an option from a closed set, a score or a probability — together with its confidence. There are no paragraphs nobody will read and no output tokens to pay for.
The Problem It Targets: Decisions Inside Software
Today, when you want a model to classify a ticket or decide whether an alert is serious, you ask for JSON, parse it and hope the format holds; when it does not, you retry and pay again. That plumbing is brittle and, at volume, expensive. A model built only to decide promises to remove those steps: the type is guaranteed by construction and the cost is a few input tokens.
Who Is Behind It: TypeSafe, Two Years in Stealth and $40M From DCVC
TypeSafe AI came out of two years in stealth on September 15, 2026, with a seed round of roughly $40 million led by DCVC. It is a San Francisco company founded in 2024, and Jev is its first System One Model, the opening release in a category the company defines as fast, automatic tasks with no chain of thought.
Read also
What System One Means
The name points to the fast, automatic answer as opposed to step-by-step reasoning. Jev does not explain why it decided: it gives a decision and a confidence value. That is exactly what makes it fast and also what makes it useless for reasoning.
The Three Primitives: Choice, Score and Noul
Every question to Jev takes one of three shapes, and none of them returns free text:
- Choice: picking one option from a closed set, with up to 255 options per question.
- Score: returning a score on a scale you define.
- Noul: returning a probability, nothing more.
Each answer comes with a confidence value, except Noul, which returns only its probability. You can also send several typed questions in a single call, so a batch of classifications costs one request.
What a Calibrated Probability Actually Means
This is the distinction worth understanding before trusting the numbers: a model saying 80% confidence does not mean it is right 80% of the time. It means exactly that when it is calibrated, that is, when the percentage matches the real hit rate. TypeSafe calls its training method Reinforcement Learning for Calibrated Decisions, and the promise is that you can use that number as a threshold inside your own code.
How You Use It From Code
What You Get Back: A Typed Field, Probability and Confidence
You send the state in plain language alongside one or more typed questions, and you get back a field with the correct type, its probability and its confidence. No parsing, no format retries and no hand-written schema validation. Model identifiers look like typesafe/jev-1.13, with a jev-latest style alias, the context window is 32,000 tokens, and there are Python and JavaScript SDKs in addition to the REST API.
Pairing It With an LLM: Who Decides and Who Writes
The intended use is not to replace anything but to split the work. Jev decides — classifies, routes, scores, verifies — and a language model writes, summarizes or explains when text is needed. In a typical flow the cheap filter runs first and the expensive model only fires when the decision justifies it, which changes the bill without changing the product.
Price and Latency: The Numbers Behind the Pitch
- $0.042 per million input tokens ($42 per billion) and free output tokens, per its published pricing.
- Between 70 and 500 milliseconds for System One tasks, as the company states.
- In a demo reported by the trade press, one query finished in 0.114 seconds at an estimated cost of $0.00008.
The 193.6x faster and 444.6x cheaper figures are TypeSafe's own comparisons in best-case scenarios against comparable language models. They are attributed to the company, not to independent measurement, and the most defensible reading is that they moved the speed and cost frontier for structured decisions, which is not the same as having a smarter model.
The Data Point a Third Party Can Verify: Vercel
Vercel reported that within 24 hours of arriving in its AI Gateway, Jev reached nearly 13% of paid teams, more than double any previous gateway launch and more than double the penetration of the GPT-5.6 family over the same period. It is the strongest item in the file, and even so it is worth being precise about what it measures: adoption inside one specific platform on day one, not market share and not model quality.
What Is Not Proven (Read Before Migrating)
Correct Typing Is Not a Correct Answer
This is the central nuance of the launch. A guaranteed output type means the interface is reliable, not that the content is true: a classifier can return the wrong option with a perfect format and 98% confidence. On top of that, the architecture is closed, and the company claims a 0% type-error rate that it itself acknowledges is not empirical, because the type is guaranteed by construction. There is no published independent verification of the speed and cost claims, and the quality benchmarks circulating are for System One tasks, not general reasoning. Access started as early access with a waitlist and is available through Vercel AI Gateway, OpenRouter, Netlify AI Gateway and AIMLAPI, with a LangChain integration.
Where It Fits and Where It Does Not
It fits classifying and routing without burning a frontier model, scoring and verifying outputs, and controlling an agent's next step without generating words. It does not fit reasoning, writing, summarizing or explaining: that still needs a language model alongside it, and the intended use is in fact to combine them.
What It Means for Your Next Integration
The industry separating deciding from writing makes a large part of AI pipelines cheaper, and Jev is the first commercial piece to say so out loud. Before migrating anything, measure your own task with real labels: error per task, effective confidence calibration (an 80% should mean being right close to 80% of the time), end-to-end latency and total cost against your current alternative. If you want to see how these pieces fit together, we already covered the Model Context Protocol standard and compared AI code editors in 2026.
Conclusion
Jev does not replace GPT or Claude, and it does not need to: it covers the part of the pipeline where text was never required. If your AI bills are full of output tokens nobody reads, that is the gap this kind of model is aiming at. Keep the caveat in mind: a correct type guarantees the shape, not the truth.
