Jev, the AI that cuts without text and at low cost

Jev does not generate any words but makes structured decisions with a confidence score in a single pass. TypeSafe claims cost and speed gains by batching questions and lists an entry price of $0.042 per million tokens, with no output billing. Reported tests show mass processing in a few seconds, and the RLCD training aims for so-called honest probabilities. Access remains limited, and the publisher lists nine failure modes.

Entry Cost, Speed, and Practical Results

TypeSafe reports that by batching thirteen questions into a single request, the cost decreases by a factor of 12.2 and speed increases by a factor of 10, without altering the responses. The publisher displays an entry cost of $0.042 per million tokens and indicates that output is not charged, as no text is generated. At this rate, it estimates that an agent making ten decisions per second in a game would cost about $7 per hour. Reported trials illustrate these orders of magnitude. Roman claims to have evaluated 700 leads with personalized messages in about 40 seconds, for a cost close to $0.09, with performance prediction, confidence score, and reporting of inconsistencies. In another test, Tarasshyn indicates that he pointed Jev at 3 million replay events and obtained, in 40 seconds and for $2.17, the review of 3,247 sessions, detection of 132 rage clicks, 116 dead clicks, and 95 JavaScript errors, followed by the opening of 213 correction pull requests.

A Model That Decides in Parallel Rather Than Writing

Jev does not behave like a chatbot: it produces no text, which TypeSafe presents as a deliberate choice. While systems like ChatGPT, Claude, or Gemini compose a response token by token, Jev performs a single input read and returns a predefined decision along with a confidence score. The absence of sequential generation allows outputs to be processed in parallel, resulting in faster, more predictable, and potentially less costly execution. In contrast, in a traditional LLM, each token depends on the previous one, a constraint that slows down application use as long as the output remains a sentence. Jev answers all questions at once, instead of producing them one by one.

Types of Questions and the Effect of Batching

The system handles three complementary request formats. Choice selects an option from a closed set, Score assigns a value on a defined scale, and Noul provides a probability between 0 and 1 for a binary question. These formats can be freely combined in a single request, with responses returned simultaneously. Adding more questions barely influences the response time, and asking one or thirteen questions takes about the same duration. In a case of routing support tickets, where keyword rules or manual review are usually required, Jev allows for parallel evaluation of the responsible team and urgency.

Reported Large-Scale Use Cases

Beyond support, public experiments describe massive processing. DevaiahShrithan ran every AI summary from arXiv from 1993 to 2026 through Jev, repeating five questions about the claimed state of the art, the presence of code, LLM-style writing, article category, and the level of emphasis in the language. Roman describes the evaluation of 700 leads with expected performance measurement, confidence score attribution, and identification of discrepancies between profile and message, in about 40 seconds for a cost of approximately $0.09. Tarasshyn reports sorting browsing sessions with, in 40 seconds and for $2.17, the analysis of 3,247 sessions, identification of 132 rage clicks, 116 dead clicks, and 95 JavaScript errors, and the automatic opening of 213 correction requests.

Probability-Oriented Training and Recognized Limits

TypeSafe describes a specific training method, RLCD, designed to calibrate what are considered honest probabilities. This approach differs from RLHF, often used to favor responses perceived as safe, and from RLVR, trained on verifiable outputs. According to the publisher, the founder of Jev co-invented RLHF at OpenAI before designing RLCD to address this blind spot. Regarding robustness, TypeSafe specifies that Jev does not return a value outside the defined set but can make mistakes within that set. The publisher distinguishes between type safety and correctness, with only the former being absolutely guaranteed, and lists nine failure modes in its documentation. Access to Jev remains limited.