Jev -The AI Model That Doesn't Talk, and Why Developers Love It

Jev is a new AI model from TypeSafe AI, founded by ChatGPT and RLHF co-inventor Diogo Almeida. Unlike large language models, Jev outputs structured probabilities ("calibrated decisions") instead of text. This makes it up to 200x faster, roughly 100x cheaper, mathematically unable to hallucinate, and ideal for automation tasks like classification, routing, and AI guardrails. It launched in early access in September 2026.

For four years, the AI industry has optimized for one thing: sounding human. Chatbots that write, argue, code, and converse. But a new model called Jev, released this week by a startup founded by one of ChatGPT's own architects, makes a provocative bet: for most of the automation work businesses actually need, human language is the wrong output entirely.

The heartbreak behind the idea

Jev comes from TypeSafe AI, co-founded by Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and co-invented reinforcement learning from human feedback (RLHF), arguably the single technique most responsible for the current AI boom.

His diagnosis of why that boom hasn't translated into ubiquitous automation is blunt. "We have lightning in a bottle, and yet it is not useful," Almeida told TechCrunch. His conclusion after years of wrestling with it: the problem is that models optimize for human language, and while the industry has been "super good at human language for four years, it's not useful for automation because computers speak a different language."

Two years ago he left OpenAI to fix that. This week, TypeSafe released Jev, a transformer-based model that is deliberately not a large language model. It doesn't output text at all; instead it produces probabilities, which the company calls "calibrated decisions."

What makes Jev different

The core design choice is to give up strings (generated text) and return type-safe structured values instead. You define the possible outputs in advance; the model returns one of them, along with a confidence score. That single decision cascades into a set of unusual properties.

It can't hallucinate. Because users define the outputs in advance, the model cannot hallucinate; its output tokens are free, and input tokens are metered by the billion rather than the million. TypeSafe frames this as mathematically guaranteed: if the schema only allows valid answers, a type error is impossible by construction. That is a sharp contrast with how AI hallucinations behave inside conventional language models.

It's radically cheaper and faster. By TypeSafe's own published figures, input tokens run about $0.042 per million (roughly $42 per billion) with output tokens free, versus $0.20 to $10 per million input for typical LLMs plus pricier output. On speed, TypeSafe claims end-to-end response times of 70 to 500 milliseconds, which it positions as 40x to 200x faster than frontier LLMs on the tasks it targets. The reason is architectural: instead of generating one token at a time, Jev samples all outputs in parallel in a single query.

It hands you honest uncertainty. Ask a normal LLM for a confidence score and it tends to be overconfident and inconsistent. Jev is built so that a stated 90% actually means roughly 90% accuracy, which is precisely what you need to safely automate a decision.

TypeSafe calls this a "System One model," borrowing Daniel Kahneman's distinction between fast, intuitive "System 1" thinking and slow, deliberate "System 2" reasoning. Jev is intuition, not reasoning: unstructured state in, typed probabilistic decision out. Almeida describes it as a "frontier-intelligence function call." That framing sits apart from the recent race toward deeper hybrid reasoning seen in models like Claude 3.7 Sonnet.

Jev vs a typical LLM at a glance

Why developers are actually excited

Interest was immediate. The company briefly lost the ability to serve users from its API because demand was so high, and developers see Jev as most useful for software automation: a cheaper and more robust way to build intelligence into code.

The early field reports are concrete:

• Vercel ran a classifier to review commands for safety. A Vercel engineer said that after replacing OpenAI's Luna model with Jev for the task, it ran five to 18 times faster and with greater accuracy.
• Bryo AI tested email classification. Against Gemini, Gemini was slightly more accurate but 10 to 20 times more expensive; what stood out most was that Jev "is the only one that hands back a real probability which makes it ideal for automating workflows."

Beyond replacing LLMs, Jev can watch them. Because using agents to monitor other agents gets expensive fast, Almeida sees Jev deployed to track LLM agent traces and prevent jailbreaks, cheap enough to sit in the loop as a guardrail. This is a natural fit for teams building autonomous agentic AI systems that need constant oversight. Another emerging use is model routing: predicting in real time whether a workload needs an expensive model, a sorting job that would be uneconomical with an LLM but is viable at Jev's cost and speed, and a core concern of the rising discipline of agent orchestration engineering.

The mental model that clarifies its limits comes from Armin Ronacher, CTO of Earendil: the model "delegates the hallucination problem a little bit to the user." At 50% probability you treat it as a coin toss and disregard it; at 95% you act on it. The intelligence is real, but the judgment about when to trust it moves back to the developer, which is exactly what makes it composable into reliable systems.

The name is the thesis

Jev is named after William Stanley Jevons, the 19th-century economist behind the paradox that says falling costs of a commodity lead to it being used more, not less. This is the same dynamic explored in detail in how Jevons Paradox is fuelling the AI boom. The bet is that cheap intelligence gets deployed everywhere. As Almeida puts it, he expects "smart software all over the place in a way that's emergent and distributed, much more like the early internet than the mega apps people are trying to build right now."

The honest caveats

This is early, and TypeSafe is refreshingly candid about it:

• The architecture is a black box. Almeida is tight-lipped about how Jev is built, and outside observers suspect it sits on top of an open-weight LLM.
• It's trained entirely on synthetic data via a method Almeida calls reinforcement learning from calibrated decisions, which he rates as a better bet than RLHF itself. That's a strong claim the field will test over time.
• Pricing sustainability is unproven. TypeSafe openly admits it can't yet prove its prices aren't subsidized, the same question that hangs over the whole industry.
• The benchmarks are self-reported. The headline "193x faster, 444x cheaper" figures come from TypeSafe's own workflow evals, built by its own team, using OpenAI and Anthropic models as the reference answers. Independent replication is still to come.
• The category is about to get crowded. Ronacher expects competitors to spring up now that the utility is visible, noting the industry "should have seen this earlier," but cheap, subsidized LLMs removed the pressure to be creative.

Why this matters

Most AI progress has been vertical: bigger models, more reasoning, higher benchmark scores. Jev is a horizontal move. It doesn't try to be smarter than GPT or Gemini; it tries to be good enough at narrow, structured decisions while being two orders of magnitude cheaper and faster, and never lying about its confidence.

If that trade holds up under independent scrutiny, the interesting shift isn't chatbots getting replaced. It's the thousands of small "smart if-statements" that were never worth an LLM API call, the classify-route-score-extract decisions buried inside ordinary software, suddenly becoming economical to automate. That is one of the quieter forces behind the defining AI trends reshaping business in 2025. That's a quieter revolution than AGI headlines, but potentially a far more pervasive one.

Almeida himself rejects the frontier-lab framing: "The main product of frontier labs is fear or hype. I would like our main product to be intelligence, not a lab in the sense of betting on infinite wealth, or a religion, or building God in a data center."

In an industry that sells the future, that may be the most contrarian pitch of all.

Frequently asked questions

What is Jev?

Jev is a transformer-based AI model from TypeSafe AI that outputs structured probabilities instead of text. It is designed for software automation rather than chat.

Who created Jev?

It was built by TypeSafe AI, co-founded by Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and co-invented RLHF.

Is Jev a large language model (LLM)?

No. Jev is a "System One model." It does not generate text; it returns type-safe structured values with calibrated confidence scores.

Why can't Jev hallucinate?

Because valid outputs are defined in advance by a schema, the model can only return allowed values, making type errors mathematically impossible.

How much does Jev cost?

TypeSafe lists input tokens at about $0.042 per million (roughly $42 per billion), with output tokens free, versus $0.20 to $10 per million for typical LLMs.

What is Jev used for?

Classification, routing, scoring, data extraction, real-time decisions, and acting as a cheap guardrail to monitor LLM agents for jailbreaks.

Sources: TechCrunch (Tim Fernholz, 18 September 2026); TypeSafe AI launch post "Introducing System One Models & Jev" (15 September 2026); TypeSafe AI documentation.