Short Answer: What is Jev?
Jev is a new AI model from the startup TypeSafe AI, launched September 16, 2026, that does not generate text at all. Instead of chatting, it returns typed, probability-scored decisions for other software to act on, and its creators say it runs up to 200 times faster and hundreds of times cheaper than a typical large language model for the tasks it targets.
Quick Summary
- TypeSafe AI, founded by former OpenAI researcher and RLHF co-inventor Diogo Almeida, launched Jev on September 16, 2026.
- Jev does not output natural language. It returns structured, typed values with calibrated probabilities and confidence scores, such as {“billing”: 0.08, “technical”: 0.85, “sales”: 0.07}.
- TypeSafe says Jev responds in roughly 70–500 milliseconds, 40x to 200x faster than a typical LLM, and its own demo shows a 0.114-second response versus 8.566 seconds for a comparable OpenAI model.
- Pricing is dramatically lower for the workloads it targets: TypeSafe lists $0.042 per million input tokens and $0 per million output tokens, versus several dollars per million tokens for frontier chat models.
- It is aimed at software automation, not conversation — early users report using it for content classification, workflow routing, and monitoring other AI agents for misbehavior.
TypeSafe AI released Jev this week, a new kind of transformer-based model built by Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and co-invented reinforcement learning from human feedback, or RLHF. Unlike the large language models Almeida helped pioneer, Jev never produces a sentence. It exists to answer structured questions with a number, not a paragraph — and developers have responded fast enough that TypeSafe briefly lost the ability to serve users from its API.
Why Build a Model That Doesn’t Talk?
Almeida told TechCrunch that ChatGPT’s success left him unsatisfied. “We have lightning in a bottle, and yet it is not useful,” he said. “I’ve been battling that problem since then. It took me a while to come to the conclusion: The problem is we are optimizing for human language… We have been super good at human language for four years, but it’s not useful for automation because computers speak a different language.”
That thesis led him to leave OpenAI two years ago and start TypeSafe AI. “TypeSafe was founded to pursue an alternative path for AI research, focused on machine-native AI,” Almeida said in a statement reported by The Register. “I spent years working on models designed to make AI better at interacting with people. But if AI is going to fundamentally change how work gets done, people can’t be the only consumers of intelligence.”
How Does Jev Actually Work?
A developer starts with a state value — a JSON object, or something as simple as the string “My card was charged twice.” Jev is then presented with the state through a set of question primitives (TypeSafe calls them Choice, Score, and Noul), each of which returns a different type of structured response with a probability attached.
The Register’s example: a question asking Jev which of three departments should handle a customer service query might return {“billing”: 0.08, “technical”: 0.85, “sales”: 0.07}, alongside a separate confidence score. That output is useless to a person looking for an answer in conversation, but it is immediately actionable for a piece of software deciding how to route that ticket.
TypeSafe calls Jev a “System 1” model, evoking the fast, intuitive mode of thinking from behavioral psychology, as opposed to the slower, deliberate reasoning associated with chain-of-thought AI agents. Architecturally, it relies on what Almeida calls “reinforcement learning from calibrated decisions,” trained exclusively on synthetic data the company generates itself. Unlike an LLM predicting one token at a time in sequence, the System 1 architecture returns all outputs to a query at once, in parallel — which is where most of its speed advantage comes from.
How Fast and Cheap Is It, Really?
TypeSafe’s own benchmarks are dramatic. The company’s published demo shows Jev responding in 0.114 seconds, compared to 8.566 seconds for a comparable OpenAI model on the same task. TypeSafe says typical response times run 70 to 500 milliseconds, which it describes as 40x to 200x faster than conventional LLMs.
On price, TypeSafe lists $0.042 per million input tokens and $0 per million output tokens for Jev, against roughly $2 per million input tokens and $12 per million output tokens for a comparable OpenAI model it cites, and a claimed 238x cost advantage against a top-tier frontier model.
Independent developer testing points the same direction, if less extremely. Pranit Sharma, a software engineer at the agentic infrastructure company Vercel, told TechCrunch that after replacing an OpenAI model used to classify commands for safety with Jev, results came back five to 18 times faster with greater accuracy. Bryo AI CTO Nikhil Mudholkar tested Jev against Gemini for classifying business emails and found Gemini slightly more accurate but 10 to 20 times more expensive — and said Jev’s calibrated probability scores were the more interesting result: “it is the only one that hands back a real probability which makes it ideal for automating workflows.”
What Is It Actually Used For?
TypeSafe is pitching Jev less as an LLM replacement than as a companion to one. Because its outputs are typed and probability-scored rather than open-ended text, backers argue it sidesteps the hallucination problem in a specific, narrow sense: since the set of possible answers is defined in advance, Jev cannot generate a fabricated fact the way a language model can. The Register pushes back a little on how clean that framing is — a structured, probability-scored answer can still be wrong, it just can’t be a fabricated citation or invented fact the way free-form text can.
Beyond classification and routing, Armin Ronacher, CTO of the open-source model harness Pi at Earendil, sees Jev as a cheap way to police other AI systems: using one language-model agent to monitor another for misbehavior gets expensive fast, but a low-cost, low-latency model built purely to output a confidence score is well suited to watching an agent’s action trace for jailbreak attempts. Ronacher also floated model routing as a use case — using Jev to cheaply predict which specific model a given task actually needs, rather than paying LLM prices just to make that triage decision. As a proof of concept, TypeSafe has also shown Jev playing the 1993 video game Doom, fed structured descriptions of the game state instead of pixels.
Why This Matters
Almeida named the model after 19th-century economist William Stanley Jevons, whose namesake paradox describes how making a resource cheaper to use can increase total consumption of it rather than reduce it. The bet embedded in Jev is that cheaper, faster machine-native intelligence won’t just replace some existing LLM calls, it will unlock a wave of “smart software” uses that were never worth paying LLM prices for in the first place — small, constant, background decisions baked into ordinary code rather than big chat-based interactions.
It is also a useful reminder that “AI model” doesn’t have to mean chatbot. Most of this year’s AI news, on this site included, has been about models that talk, reason out loud, or act as agents. Jev is a bet that a meaningful share of AI’s real economic value will come from the opposite: models that never produce a sentence at all, built to be invisible infrastructure inside someone else’s software rather than a product people talk to directly.
Frequently Asked Questions
Is Jev a large language model?
No. TypeSafe describes Jev as a transformer-based model, but unlike an LLM it does not generate natural-language text. It returns typed, structured values with calibrated probabilities and confidence scores instead of words.
What is Jev used for?
Early users are applying it to software automation tasks such as classifying customer support tickets, routing requests between departments, screening commands for safety, and monitoring other AI agents for misbehavior or jailbreak attempts.
Is it true that Jev cannot hallucinate?
TypeSafe markets Jev as hallucination-free because its possible outputs are defined in advance by the developer, so it cannot fabricate an open-ended fact or citation the way a language model can. That does not mean its probability-scored answers are always correct, only that a wrong answer takes a different, more constrained form.
Who created Jev?
Jev was built by TypeSafe AI, a startup founded by Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and co-invented reinforcement learning from human feedback (RLHF). He left OpenAI roughly two years before Jev’s September 16, 2026 launch.
How much does Jev cost compared to a typical LLM?
TypeSafe lists pricing of $0.042 per million input tokens and $0 per million output tokens for Jev, which the company says is roughly 238 times cheaper than a top-tier frontier model for the narrow, structured tasks Jev targets. It is not a general-purpose substitute for conversational AI pricing comparisons.
Sources: TechCrunch, The Register.
About the author
Kai Williams
Kai Williams has been in marketing for years, with a long background in SEO before AEO had a name. He stepped into Answer Engine Optimization the moment AI started reshaping how people search, and has been tracking the shift ever since. At Prompt Insider, he covers AEO, AI marketing, and the future of search, breaking down what is changing and what brands need to do about it.
