Former OpenAI researcher redefines large models: no more word-by-word generation, up to 200x faster, costs as low as 1/400 Former OpenAI researcher and one of the co-inventors of ChatGPT, Diogo Almeida, has announced the launch of a new model called Jev. Unlike traditional models that first generate natural language or JSON and then rely on software to parse it, Jev directly outputs structured decisions and probabilities, targeting tasks like fraud detection, workflow routing, and tool invocation. According to official reports, Jev achieves a speed boost of 20 to 200 times, reduces costs by 40 to 400 times, offers free token output, and is optimized for composable decision-making intelligence. Its new training method, RLCD, focuses on enhancing the model's probability calibration capabilities, with a cost of just $0.042 per million input tokens and latency of around 150 milliseconds.