Former OpenAI researcher Diogo Almeida has taken TypeSafe AI out of stealth with Jev, the startup’s first “System One Model” designed to make fast, structured decisions directly inside software.
Jev is not trying to become another chatbot. Instead of generating long, open-ended answers, the model is built around predefined choices that software developers set in advance. It then evaluates incoming information, selects from those available options and returns confidence scores alongside its answer.
TypeSafe says this narrower approach allows Jev to run much faster and at a far lower cost than general-purpose large language models. The company is pitching the system as something closer to an intelligent software function than an AI coworker.
Jev Gives Software Decisions Instead of Generated Text
Most large language models generate text one token at a time. That flexibility allows systems such as ChatGPT, Claude and Gemini to write articles, answer questions, create code and hold long conversations. It also creates unpredictability because the model has far more freedom over what it produces.
Jev takes a different route. Developers define the possible structure of an answer before the model is called. Jev then analyzes the information it receives and chooses from those predefined options, returning a structured response that the rest of the application can immediately use.
TypeSafe describes the approach as a “frontier-intelligence function call.” The idea is fairly simple: software sends Jev information, Jev makes a constrained judgment, and the application receives something usable without needing to process a long piece of generated text.
That makes the model more suited to quick decisions inside software than open-ended creative tasks.
TypeSafe Says Jev Costs $42 Per Billion Input Tokens
Pricing is one of the most aggressive parts of TypeSafe’s launch.
The company lists Jev at $0.042 per million input tokens, which works out to roughly $42 per billion input tokens. TypeSafe also says output is currently free because Jev does not generate lengthy token-by-token responses in the same way conventional language models do.
That pricing structure could matter in applications where AI has to run repeatedly in the background. A single chatbot request may not cost much, but thousands or millions of model calls inside software can quickly become expensive.
TypeSafe believes Jev changes that equation. By making individual AI decisions much cheaper, the company wants developers to use intelligence in parts of software where calling a conventional frontier model would previously have been too costly.
Jev Responses Can Arrive in 70 to 500 Milliseconds
Speed is another major part of the Jev pitch.
TypeSafe says the model can return responses in roughly 70 milliseconds to 500 milliseconds, depending on the task. The company also claims its System One approach can operate around 40 to 200 times faster than certain workflows built around traditional large language models.
That kind of latency could make a noticeable difference inside applications. Waiting several seconds for a chatbot to respond is usually acceptable because the user expects a conversation. Waiting several seconds every time software needs to classify a request, score a record or decide where information should be routed is much harder to justify.
Sub-second inference gives developers more room to place AI inside ordinary application logic without making the entire product feel slower.
Why TypeSafe Says Jev Cannot Hallucinate
TypeSafe makes one of its boldest claims around hallucinations, saying Jev “can’t hallucinate.”
The reasoning comes from the way the model is designed. Jev does not freely generate arbitrary text. Developers define the available answer types beforehand, which means the system cannot suddenly produce an unexpected structure or invent an entirely different format.
That does not mean Jev will always make the correct decision. It can still select the wrong predefined answer. The difference is that the system stays inside the boundaries developers give it instead of producing unrestricted text that may contain fabricated details.
For production software, that distinction matters. A model that occasionally chooses the wrong option still needs monitoring, but predictable outputs are much easier for applications to validate and control.
Confidence Scores Are Built Into Every Decision
Jev also returns confidence information alongside the decisions it makes.
TypeSafe says the model was trained using an approach it calls Reinforcement Learning for Calibrated Decisions, or RLCD. The goal is to make the model not only produce a decision but also communicate how confident it is in that result.
That confidence information gives developers more control over what happens next. An application could automatically accept a highly confident decision, send a lower-confidence case to another model or route uncertain results to a human reviewer.
This turns Jev into more than a basic classifier. The system can become one stage inside a larger decision-making pipeline, with different actions triggered depending on how certain the model appears to be.
Jev Could Work Like a Smart If-Statement Inside Software
TypeSafe sees Jev being used for the kinds of small decisions that applications make constantly.
The model could help sort incoming requests, classify records, score database entries, choose between workflows or decide where a task should be routed. These are not necessarily jobs that require a full conversational model, but they can still benefit from AI-level judgment.
Jev could also sit around other AI systems rather than replace them. TypeSafe says the model could review prompts, inspect outputs or help identify jailbreak attempts before a larger language model continues with another action.
That creates an interesting division of labor. A generative model could handle complex reasoning or content creation while Jev manages fast checks and structured decisions around it.
TypeSafe Built Jev After Two Years in Stealth
TypeSafe says it spent roughly two years developing the technology behind Jev before publicly launching the company.
Founder Diogo Almeida previously worked at OpenAI and contributed to research related to instruction-following language models. His work at TypeSafe appears to focus on a different question: whether every AI task really needs a large generative model.
The company’s System One concept is built around smaller, faster decisions rather than broad conversational intelligence. TypeSafe says its technology stack includes a new model architecture, a parallel sampler and its RLCD training method.
Jev is currently being introduced through early access as the company begins putting the system in front of more developers.
The Name Jev Comes From Jevons Paradox
The name Jev is a reference to economist William Stanley Jevons and the idea commonly known as Jevons paradox.
The paradox describes how making something more efficient can sometimes increase overall consumption instead of reducing it. When a resource becomes cheaper or easier to use, people may simply start using much more of it.
TypeSafe believes AI could follow a similar pattern. If software intelligence becomes dramatically cheaper and faster, developers may not just save money on existing AI workloads. They may start adding models to parts of applications that previously had no AI at all.
That could mean a much larger number of smaller model calls running behind the scenes rather than fewer expensive interactions with one large model.
TypeSafe’s Performance Claims Still Need Wider Testing
The early numbers around Jev are impressive, but they remain company-reported results.
TypeSafe says some of its internal workflow evaluations showed Jev operating substantially faster and cheaper than comparison systems. The company has also acknowledged that some of those results may represent the upper end of what developers should expect depending on the workload.
Independent testing will be important as Jev reaches more users. Real-world accuracy, confidence calibration, latency under heavier traffic and performance across unfamiliar tasks will all matter more once developers begin running the model outside TypeSafe’s own evaluation environment.
The pricing and speed are enough to make Jev worth watching, but broader evidence will determine whether those advantages hold up in production.
Jev Is Not Trying to Replace Large Language Models
Jev makes more sense when it is viewed as a different category of AI rather than another competitor to GPT, Claude or Gemini.
Large language models are designed for flexibility. They can answer questions, reason across long contexts, generate code and produce open-ended content. Jev intentionally gives up much of that flexibility in exchange for speed, lower cost and predictable outputs.
That trade-off could make it useful as a background layer inside software.
If TypeSafe can maintain the latency, pricing and reliability it is promising, Jev could become part of an emerging class of AI systems that users rarely interact with directly but applications call constantly.
You may never open a chat window with Jev. The software you use could still be asking it questions all day.
Sources
TypeSafe AI — Introducing System One Models and Jev
https://typesafe.ai/blog/introducing-system-one-models-and-jev
