ChatGPT co-developers have created 'Jev,' an AI model that processes tasks using a different method than LLM, enabling low-cost and ultra-fast task execution with GPT-5.6 Terra-level performance.



AI company

TypeSafe AI announced ' System One Models ,' an AI that solves tasks using an approach different from large-scale language models (LLMs), and ' Jev ,' the first model in the System One Models series, on September 15, 2026, local time. TypeSafe AI is led by CEO Diogo Almeida, an AI researcher who was involved in the development of ChatGPT, and claims to have achieved high-speed and low-cost AI by using a fundamentally different method from LLMs.

Introducing System One Models & Jev - TypeSafe AI Blog
https://typesafe.ai/blog/introducing-system-one-models-and-jev




Cutting-edge models from OpenAI and Anthropic can handle a wide variety of operations, including coding and modeling. However, the reality of these cutting-edge models is that they 'continue to generate the rest of the text,' and this mechanism itself is a bottleneck in terms of efficiency and speed. Therefore, TypeSafe AI has developed 'System One Models' as an AI model with a different approach from LLM, and has announced 'Jev' as its first model.

The video included in Almeida's post below clearly shows the difference in operation between Jev (left) and a typical LLM (right). A typical LLM generates text sequentially from left to right, and that text ultimately becomes a deliverable such as a 'novel,' 'code,' or 'specification document.' On the other hand, Jev does not aim to generate text in the first place, but outputs the deliverable all at once according to a predefined structure. It is also characterized by its extremely fast processing speed.




Jev is also characterized by its high cost efficiency. The fee per 1 million tokens input is $0.042 (approximately 6.5 yen), and the output is listed as '$0' or 'unmeasurable due to extremely low cost.'

The graph below shows the cost per task on the horizontal axis and the task completion accuracy on the vertical axis. Jev performs tasks with the same quality as GPT-5.6 Terra, but at about 1/100th of the cost.



Jev's output and structure are predefined, and it is said that no errors will occur. In addition, all outputs come with a confidence score. The graph below shows the error rate of tool calls in each model, with Jev having an error rate of 0%.



The following are the results of a comparison of the execution speed of the task 'following links on Wikipedia to reach a target page' using Jev and a third-party model. All models were run in high-speed non-inference mode, but Jev was still faster.




A video of Jev playing DOOM has also been released. The cost of one hour of gameplay was $7 (approximately 1090 yen).




TypeSafe AI has begun accepting applications for its Early Access program at the same time as Jev's announcement. Additionally, documentation for System One Models is available at the following link.

Introduction - TypeSafe AI
https://docs.typesafe.ai/introduction

Related Posts:

in AI, Posted by log1o_hf