In 2026, a new category of artificial intelligence emerged that wasn't designed for conversation. While the market was vying for the best text-writing model, a startup called TypeSafe introduced JEV, a model designed for a different function: decision-making. It doesn't write emails, summarize reports, or engage in casual conversation. It receives a situation, evaluates the possible options, and returns a choice, along with the probability of it being correct.
The response was swift. In the first 24 hours, nearly 13% of Vercel's AI Gateway paying teams tested the model, the highest adoption rate for a launch ever recorded by the platform, according to Vercel itself. For those working in retail and technology, JEV signals a trend worth following: AI is beginning to divide between models that produce content and models that make operational decisions at scale.
On this page you will find:
JEV in 6 basic points:
- JEV is an AI model from TypeSafe that does not generate text. It receives a context and returns a typed decision, with the probability of each option;
- It was created to be called by code., within software, and not for dialoguing with people;
- It responds in three formats. (yes or no, choose between categories and rate on a scale) always with a degree of confidence;
- Latency ranges from 70 to 500 milliseconds per call., according to TypeSafe;
- The announced price is US$0,042 per million entry tokens., with no charge for leaving;
- It serves to classify, route, prioritize, and verify tasks. which currently consume the time of customer service, marketing, and operations teams.
Who created JEV and why did its launch generate so much buzz?
JEV was presented by Diogo Almeida , founder of TypeSafe and former researcher at OpenAI . He participated in the work surrounding InstructGPT , one of the foundations that gave rise to ChatGPT ( which is why some of the press described him as a co-creator of the assistant). The company operated in closed mode for about two years before opening the model to the public in September 2026, and announced an initial round of US$40 million led by DCVC , according to InfoMoney.
The name also reveals the project's intention. TypeSafe calls its line " System One Models ," a reference to the quick and intuitive reasoning described by Daniel Kahneman in Thinking, Fast and Slow . "Jev" is a tribute to the economist William Stanley Jevons . The idea behind the brand is straightforward: a model for quick and repetitive judgments, not for lengthy reflections.
How JEV works
The logic of a JEV (Job Value Entity) is different from that of a chatbot. Instead of an open conversation, the developer sends two things: a state, which is the context of the problem in text or JSON, and a typed question, with the valid options already declared. The model responds only within those options.
There are three types of responses. The first is yes or no, useful for validation (for example, whether a message contains enough data to proceed). The second is a choice between categories, with a limit of up to 255 alternatives per question. The third is a rating on a scale defined by the person asking . In all cases, each option returns its probability.
Two technical details support the proposal. TypeSafe states that the guarantee is structural, not statistical: the format of the response is always valid, but the accuracy of the choice is not guaranteed. And the model is trained with a method the company calls RLCD ( reinforcement learning for calibrated decisions ), aimed at aligning informed confidence with the actual success rate; the goal is that 80% confidence corresponds to being correct around 80% of the time.
| Criterion | IS IN | Traditional LLM |
|---|---|---|
| Main goal | Deciding between options | Generate text or code |
| Output | Typed option + probability | Text in natural language |
| Who makes the call? | The system code | A person, in general |
| Typical latency | 70 to 500 ms | Seconds |
This division of roles summarizes the recommended use: generative AI remains the choice for writing, translating, and summarizing; JEV comes in when the software needs a quick, standardized decision to follow on its own.
Why retail and technology should watch the JEV.
The promise of JEV targets precisely the decisions that are repeated thousands of times a day in a digital operation. In customer service , such a model can read a call and determine which area to forward it to, or classify the urgency before a human opens the message. In marketing and sales , it can score a lead and indicate whether it continues through automation or goes directly to a salesperson. In the catalog , it can check if an AI-generated description contradicts the product data before the page goes live.
This last point explains why the launch caught the attention of those already using generative AI . One of the uses cited by TypeSafe is to verify the output of another model —to flag, for example, a text that promises something out of register or that contradicts the previous conversation. Restricting the response to a closed set of options reduces the space for the model to invent information , although it does not eliminate the chance of a wrong choice: an error can still misroute a call or set the wrong priority.
Cost helps to understand market appetite. Frontier models charge from $0,20 to $10 per million tokens for entry; JEV was announced at $0,042, with no charge for exit. In the company's own tests, the model was approximately 194 times faster and 445 times cheaper than language models in certain flows. The number is impressive, but calls for editorial caution: the benchmark for these tests was the average of the responses from other AI models, that is, it measures agreement, not accuracy confirmed by independent evaluation.
It's still early. The JEV has only been open for a few weeks, and much of the performance evidence comes from the manufacturer itself. Even so, the direction of this movement is of interest to digital retail: some tasks that currently depend on manual rules or an expensive and slow LLM could migrate to specialized decision models. It's worth observing, testing on a small scale, and measuring with your own data before making any big bets.
Conclusion
JEV doesn't replace ChatGPT or your team. It represents a new facet of artificial intelligence (models that decide instead of writing), and this facet is increasingly appearing in e-commerce, customer service, and marketing tools. Understanding the difference now avoids confusion later when choosing the right technology for each task. To continue following how AI is changing digital retail operations, explore other content about artificial intelligence and technology here on the e-Plus blog.
FAQ
What is JEV?
JEV is an artificial intelligence model from the startup TypeSafe that doesn't generate text, but rather decisions. It receives a context and a set of options and returns the most likely choice, with a degree of confidence. It was created to be used within software, via code, in tasks such as classifying and routing information.
What is the difference between JEV and ChatGPT?
The difference lies in their function. ChatGPT is a language model that writes text, code, and conversational responses. JEV decides between predefined options and returns one of them with a probability of being correct. One serves to generate content; the other, to make quick and standardized decisions in systems.
What is the purpose of JEV in e-commerce?
JEV is used to automate repetitive decisions in a digital operation. Possible uses include routing support requests to the correct department, prioritizing tickets, scoring leads, and verifying that AI-generated content is consistent with product data before publication.
Does JEV eliminate AI errors and hallucinations?
It doesn't eliminate it. Restricting responses to stated options reduces the scope for the model to invent information, which decreases hallucinations. But the guarantee is structural, not about accuracy: the JEV can still choose the wrong option. That's why it provides a confidence level, which helps the system decide when to act alone or call a human.
