TypeSafe AI has released Jev, a new transformer-based model that does not generate paragraphs, code or chatbot replies. Instead, it evaluates a piece of information and returns a structured decision, along with probabilities that software can use.

The San Francisco-based startup was founded by Diogo Almeida, a former OpenAI researcher and co-inventor of reinforcement learning from human feedback (RLHF), alongside Erik Gafni and Sasha Sheng. It emerged from stealth on September 15 with a $40 million seed round led by DCVC. Jev is available in early access through a waitlist.

A model for the decisions inside software

An online service may not need an AI model to write a response every time it receives a request. It may need to decide whether a support ticket is about billing, whether an email is urgent, whether a command is safe or which model should handle a task.

These are relatively small decisions, but they appear throughout an automated workflow. Developers can write rules for simple cases. For more ambiguous inputs, they may use an LLM and ask it to return JSON. That approach can work, but the application still has to parse the model’s response, validate its structure and deal with an answer that may be wrong or incomplete.

Jev takes a different approach. A developer gives it a state—such as a customer message, a document or the current state of a game—and defines the kind of answer the application needs. Jev returns a typed value rather than open-ended text.

TypeSafe calls Jev a “System One model”, a name inspired by Daniel Kahneman’s distinction between fast, intuitive System 1 thinking and slower, deliberate System 2 reasoning. In TypeSafe’s framing, Jev is meant for focused judgments that can be composed inside a larger software system, not for solving every problem in one long prompt.

Choice, Score and Noul

The model currently exposes three question types, or primitives:

• Choice selects one option from a defined list, such as billing, technical support or account access. It also returns the probability for each option.

• Score places an input on a developer-defined scale, such as a bug’s severity or a customer’s level of frustration.

• Noul evaluates a yes-or-no statement and returns a probability between zero and one.

Several questions can be sent with the same state and evaluated in parallel. A support system could, for example, classify a message, check whether it requests a refund and estimate its urgency in one request. The application can then combine those results with ordinary code and decide whether to automate the next step or send the case to a person.

This design also changes what “hallucination” means. Because the developer defines the possible output types, Jev cannot invent a new category or return malformed JSON in the way a text-generating model might. But a constrained answer is not necessarily a correct answer. The model can still choose the wrong option, and a probability is not an explanation or a guarantee. TypeSafe’s documentation also notes that calibration describes the behaviour of groups of predictions; it does not guarantee that any individual prediction is right.

Faster and cheaper, according to TypeSafe

TypeSafe says Jev produces responses in roughly 70 to 500 milliseconds. Its pricing is listed at $0.042 per million input tokens, or $42 per billion tokens, while output tokens are free. The company compares that with frontier LLMs that generate output sequentially and says Jev can be 40 to 200 times faster for comparable System One-shaped tasks.

Those are the company’s figures, and the comparison depends on the task, the competing model and the way each workflow is measured. Jev’s advantage is not that it can replace a general-purpose LLM in every situation. It is that a short, bounded decision can be produced without generating a long response one token at a time.

TypeSafe says Jev uses a parallel sampler and is trained with a method it calls Reinforcement Learning for Calibrated Decisions, or RLCD. The company says the model is trained exclusively on synthetic data. It has not publicly disclosed the model’s parameter count, weights or full architecture.

Where Jev could fit

TypeSafe is positioning Jev as a layer inside applications rather than as a chatbot. Potential uses include:

• routing customer requests and support tickets;

• classifying documents, messages or code;

• scoring risk, urgency or quality;

• choosing which model should handle a request;

• checking an AI agent’s proposed action; and

• making frequent decisions in games, simulations and other real-time software.

The company’s launch demonstrations include Jev controlling a game such as Doom by receiving structured information about the game state and selecting its next action. The point is not that Jev is a better game-playing system than a hand-coded bot. It is that a model can make repeated decisions at a latency and cost that would be difficult to justify with a large conversational model.

Jev can also work alongside an LLM. One model could understand a user’s request and produce a response, while Jev checks whether the request belongs to an approved category, whether a proposed action looks risky or whether the task should be escalated. In this setup, code remains responsible for permissions and hard policies; Jev supplies a probabilistic signal for the parts that are difficult to capture with fixed rules.

A bet on machine-to-machine AI

TypeSafe’s larger argument is that much of the next phase of AI will happen without a person reading every model response. AI agents will call other models, software will evaluate incoming data and applications will make small decisions continuously. Those interactions need outputs that are fast, inspectable and easy to connect to code.

The company has named Jev after economist William Stanley Jevons, whose work is associated with the Jevons paradox: making a resource more efficient or cheaper can increase its overall use. TypeSafe believes the same could happen with machine intelligence. If a useful decision costs almost nothing and arrives in milliseconds, developers may put it into many more parts of a product.

That makes Jev less of a challenger to ChatGPT-style assistants than a specialised partner for them. LLMs remain useful when a system needs language, explanation, planning or code. Jev is aimed at the quieter layer underneath: the repeated choices that tell software what to do next.

Whether TypeSafe’s performance claims hold up across independent tests remains to be seen. But Jev points to an important shift in AI design. The most useful model for a particular job may not be the one that says the most. It may be the one that returns a clear, bounded signal exactly when software needs to make a call.