TypeSafe AI's Jev doesn’t write, code or chat. So why are developers obsessed with it?

The newest AI model getting developers excited is not trying to be the next ChatGPT. It does not write emails, generate code or hold a conversation. It does not even try to explain why it made a particular decision.

Instead, Jev, a model from TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, is designed to make decisions.

Give it some unstructured information and a set of questions, and Jev returns structured choices, scores or probabilities that an application can act on directly.


That has caught the attention of developers because a lot of the work AI systems do behind the scenes does not actually require a chatbot-style answer.

That is the bet behind Jev.

Also Read: Zoho brings AI to everyday work with Zia Chat, adds seven vertical SaaS offerings

So, what exactly is Jev?

Think about a customer sending a message to an online support system: “I was charged twice for my subscription. Can you refund me?”

A conventional large language model could read the message and return something like: This appears to be a billing issue, so the request should be routed to the billing team.

But the software does not really need that sentence. It needs something closer to:

Billing: 0.94

Technical support: 0.04

Sales: 0.02

That is the kind of job Jev is designed to do.

TypeSafe describes Jev as a “System One” model, borrowing from psychologist Daniel Kahneman’s distinction between fast, intuitive thinking and slower, deliberate reasoning. The company says its System One models are built for fast, structured decisions inside software.

Jev currently supports three broad types of decisions: choosing from predefined options, scoring something against a scale and estimating whether something is true.

The important difference is that its outputs are structured from the start. Developers do not have to ask a language model for an answer, extract the useful piece from a paragraph and then convert it into something their software can understand.

TypeSafe calls this a “frontier-intelligence function call”: unstructured information goes in and typed, probabilistic decisions come out.

Why does that matter?

Because AI agents make a lot of small decisions.

An agent might use a large language model to reason through a complicated task. But around that reasoning are dozens of smaller questions:

• Which agent should handle this request?
• Should this tool call be allowed?
• Is the answer good enough to send?
• Should the workflow continue or stop?
• Does this task need a more powerful model?
• Should a human take over?

Using a frontier language model for every one of these decisions can make an AI system slower and expensive.

Jev is effectively trying to become the cheaper decision-making layer around those larger models.

Developers are already testing it this way. Vercel engineer Pranit Sharma said on X that replacing a conventional language model with Jev for command-safety classification made responses five to 18 times faster in his test, while also improving accuracy.

Vercel said in a September 18 blog post that Jev became the fastest-adopted model in its AI Gateway history. Within 24 hours of launch, nearly 13% of paid teams were using it, according to the company.

Also Read: OpenAI expands GPT-6 lineup with cheaper Sol and Luna models

Jev is cheap. Very cheap.

TypeSafe lists Jev at $0.042 per million input tokens, with no separate charge for output tokens. The company says the model typically responds in roughly 70 to 500 milliseconds for the workloads it targets.

The economics matter because software applications can make enormous numbers of small decisions. If an AI system has to classify 100,000 customer messages, evaluate thousands of agent actions or route millions of requests, paying frontier-model prices for every decision can quickly become a problem.

This is also where Jev gets its name. According to TypeSafe, it is a reference to economist William Stanley Jevons and the Jevons paradox, which describes how improvements in efficiency can lead to greater consumption of a resource.

Fast and structured does not mean correct

TypeSafe says Jev “can’t hallucinate”, but that claim needs some unpacking.

If a developer gives Jev five categories, it cannot suddenly invent a sixth category in its response. But that does not mean it cannot make the wrong decision. Jev can still select the wrong category or give a high probability to an incorrect answer.

An incorrect customer-service classification may be annoying. An incorrect decision about whether an AI agent can execute a destructive command is a very different problem.

A September 21 test reported by VentureBeat showed that risk. In an experiment by an Octomind engineer involving the command rm -rf ~/.ssh, Jev initially gave a 0.76 probability that the command should be blocked. After adversarial information was inserted into the input, the block probability fell to 0.48. VentureBeat noted that the test involved a single command in one integration, so it does not establish how often such manipulation would work.

TypeSafe itself warns that adversarial content can move Jev's answer, while Pydantic has cautioned that Jev-based guards should sit alongside deterministic checks rather than replace them.

So is Jev really doing something new?

Large language models can already return structured outputs. OpenAI, for example, introduced Structured Outputs in 2024, allowing developers to require responses to conform to a predefined schema.

But Jev is designed around the decision itself. Its answers are returned as typed choices, scores or probabilities rather than a block of generated text that software has to interpret. TypeSafe says Jev evaluates declared questions in parallel rather than generating a sequence of tokens one after another.