As MarktechPost reports, the 2022 ChatGPT moment taught artificial intelligence how to communicate with humans. One of its creators now believes the next step is AI that interacts with software rather than humans. TypeSafe AI released Jev. Jev is based on Transformers, but is not a large language model. It does not generate text. Instead, you send state and typed questions, and it returns typed decisions with probabilities on which the code can make decisions. The model is available as a hosted API in early access, but TypeSafe has not published the weights, number of parameters, or self-hosting capability. What is the System One model? The name is borrowed from Daniel Kahneman, who distinguished between intuitive and slow reasoning. The TypeSafe team claims that models tuned by RLHF to human preferences have led to the creation of chat, as well as overconfidence and mode failures. These shortcomings require human participation in the process. Jev uses a new stack: a new architecture, parallel selection and Reinforcement Learning for Calibrated Decisions (RLCD). TypeSafe
did not disclose the architecture. How does the Jev API work? One endpoint handles all requests: POST https://api.typesafe.ai/v1/systemone. The body conveys the state, model and question map. The documentation defines three types of questions: Primitive, Asks, Returns, Choice, Score, Noul. Questions are executed in parallel and in isolation based on the same state. TypeSafe claims that adding questions has little effect on response time. Choice supports up to 255 options. Each Choice and Score response has a confidence value between 0 and 1, which TypeSafe derives from the shape of the probability distribution. In the example from the documentation, billing gets 0.84, but confidence only gets 0.596, since technical still holds 0.159. The documentation suggests three options: act on high confidence, revise the averages, and send low confidence to the person. Thresholds should be scaled based on the cost of an incorrect action. Jev is priced at $42 per billion input tokens, while TypeSafe prices existing LLMs at $0.20 to $10 per 1 million input tokens. In the recorded demo video, Jev completed the task in 0.
114 seconds for $0.000081, while GPT-5.6 Terra took 8.566 seconds for $0.013880. The TypeSafe team claims that Jev is 193.6 times faster and 444.6 times cheaper. These numbers are based on TypeSafe's own workflow assessments. The community reportedly began developing projects using Jev just days after its launch. For example, Vercel CEO Guillermo Rauf reported that Jev was 18 times faster at the 95th percentile compared to GPT Luna and more accurate. His post also noted that the fx reviewer was still working on Luna. Other examples include: Bryo AI, where CTO Nikhil Mudholkar found Gemini to be slightly more accurate but 10-20 times more expensive; and browser-based agents such as jev-ultrafast, which searched for flights from Zurich to London in 7.1 seconds. Jev was also used to drive Uber on a real Android device and completed the first StarCraft combat mission. It's important to note that Jev produces typed solutions with probabilities rather than strings. Entry tokens cost $0.042 per 1 million and output tokens are free. TypeSafe reports response times between 70 and 500 milliseconds. Basic
e benchmarks are carried out by the supplier, and it is recommended to test on your own data. The material was prepared by a neural network based on the original publication. AI may make mistakes in details (numbers, names, dates, quotes) - for important information, check with the original source. Google Gemini made the first hacker attacks on the company OpenClaw released update 2026.9.5 with atomic updates and new features Kyutai released the Voice of Reason model for verbal solving of mathematical problems. Nokia opens access to AnyJev, a library for decision-making without training. OpenAI introduced GPT-6 Sol and Luna with new API pricing, reducing the cost by 50%.