News

Jev Does Not Want to Talk to You. It Wants to Make Decisions for Your Software.

• By John K. Waters
• 09/23/2026

The generative AI industry has spent years teaching machines to converse with people. TypeSafe AI is betting that the next opportunity lies in building models that communicate directly with software.

The San Francisco startup has released Jev, an artificial intelligence model designed to make fast, structured decisions inside applications. Unlike ChatGPT, Claude, or Gemini, Jev does not generate prose, write code, or explain its reasoning. It returns predefined values accompanied by probabilities that software can use without parsing a natural-language response.

TypeSafe calls Jev the first “System One” model, borrowing from the distinction between fast and deliberative thought popularized by psychologist Daniel Kahneman. The company’s larger argument is that language models are optimized for communicating with humans, while much software automation requires smaller, more predictable judgments.

“Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out,” TypeSafe founder Diogo Almeida wrote in the company’s September 15 announcement.

The model arrives with aggressive claims about speed, price, and reliability. Most of those claims are based on TypeSafe’s own evaluations, however, and have not been independently reproduced.

Intelligence without the conversation
A company could give Jev a customer-service message and ask which department should receive it. The developer would define the allowable answers, such as billing or technical support, before sending the request.

Jev could also estimate the customer’s level of frustration or calculate the probability that the message contains a refund request. Application code could then decide whether to act automatically or send the case to an employee.

TypeSafe’s developer documentation divides those judgments into three primitives. Choice selects from a predefined set of options. Score evaluates an input against ordered levels, while Noul returns a probability that a statement is true.

The strange name “Noul” combines “no” and “null,” according to the company’s documentation. In practical terms, it functions as a probabilistic yes-or-no question.

The possible outputs are established in advance. That means Jev cannot return an unexpected category or malformed response outside the requested structure.

TypeSafe describes this property as an inability to hallucinate. That claim requires qualification. Jev may be unable to invent an answer outside the supplied schema, but it can still select the wrong permitted answer. Structural reliability is not the same as factual accuracy.

Putting uncertainty into the output
TypeSafe’s more consequential claim concerns calibration.

Jev is trained using a method the company calls Reinforcement Learning for Calibrated Decisions, or RLCD. Instead of optimizing the model to produce text preferred by human reviewers, TypeSafe says it trains Jev so its probability estimates more closely match observed outcomes.

A well-calibrated model that assigns 80 percent probability to a large group of comparable predictions should be correct about 80 percent of the time. That does not mean any individual prediction carrying an 80 percent probability will be correct.

TypeSafe acknowledges that limitation in its documentation. Calibration is measured across groups of predictions and does not guarantee the accuracy of an individual answer.

For Choice and Score questions, Jev returns the full probability distribution and a confidence measurement. Developers can use that information to establish thresholds. A high-confidence, low-risk decision might be automated, while an uncertain or consequential one could be referred to a person or a more capable reasoning model.

That arrangement shifts some responsibility from the model to the application developer. Jev supplies probabilities, but the surrounding software determines what level of uncertainty is acceptable.

A narrower model with broader ambitions
TypeSafe is not presenting Jev as a replacement for general-purpose language models. The company is positioning it as a specialized layer beneath them.

An AI agent could use a large language model to create a plan or communicate with a user. Jev could handle repeated decisions within the workflow, such as choosing a tool, deciding whether an operation succeeded, or determining when to stop trying.

The approach addresses a genuine mismatch in many AI applications. Developers frequently ask language models to produce JSON or select from a fixed list, even though those models were designed to generate open-ended sequences of text. The output then must be validated before software can trust its structure.

Jev trades open-ended generation for constrained responses. It evaluates multiple questions against the same input in parallel rather than generating an answer one token at a time.

TypeSafe says this design makes Jev substantially faster and cheaper than frontier language models on comparable decision tasks. The company reports response times between 70 and 500 milliseconds. Its most favorable workflow tests showed Jev running 193.6 times faster and costing 444.6 times less than the language models used for comparison.

TypeSafe also concedes that those figures likely represent the upper end of real-world improvements. Members of its model-capabilities team created the workflows, leaving room for bias. The company used the average outputs of OpenAI and Anthropic models as its reference rather than independently established ground truth.

Jev’s current listed price is $0.042 per million input tokens. Output tokens are not charged because the system returns structured values rather than generating lengthy text.

The model accepts text, JSON objects, and text arrays. It does not currently process images or audio. Its listed context window is 64,000 tokens, subject to additional limits on the shared state and longest question.

Early developer interest
Jev has already reached some common AI development infrastructure. Vercel added the model to its AI Gateway and exposed it through an experimental evaluation API in the Vercel AI SDK.

The startup also emerged from stealth with a $40 million seed round led by DCVC. Forbes reported that the financing valued TypeSafe at $200 million.

Almeida previously worked at OpenAI and contributed to research behind instruction-following models and reinforcement learning from human feedback. He founded TypeSafe with Erik Gafni and Sasha Sheng.

The company named Jev after 19th-century economist William Stanley Jevons. Jevons paradox describes how making a resource more efficient can increase its overall consumption. TypeSafe expects cheaper machine intelligence to produce a similar expansion in demand.

That prediction may prove more important than the model’s initial benchmarks. If inexpensive probabilistic judgments become embedded throughout software, AI could become less visible while exerting more influence over routine decisions.

Whether that happens will depend on reliability, not merely speed. Companies will need evidence that Jev’s probabilities remain calibrated on their own data and under changing conditions. They will also need controls for decisions whose consequences are difficult to reverse.

Jev offers a plausible alternative to using a large language model for every AI task. It has not yet established that businesses should entrust it with every decision it can make.