Diogo Almeida presents a model at TypeSafe AI that responds in 70 to 500 milliseconds by choosing from 255 options, without text generation or parsing layers.
On September 16, 2026, Diogo Almeida, co-creator of ChatGPT and founder of TypeSafe AI, opened early developer access to Jev, a model designed for a very specific task: deciding between predefined options. Where a conversational model produces text token by token and then is read back by layers of parsing and guardrails, Jev samples in parallel and directly returns a calibrated decision, from a maximum of 255 discrete choices. His training is based on a calibrated decision-oriented reinforcement learning (RLCD) method.
The interest announced is economic and temporal. The response time is between 70 and 500 milliseconds, while conversational models take from three seconds to several minutes on comparable tasks; on workflow execution, TypeSafe claims a factor of up to 193.6 compared to GPT-6 Astra and Claude Fable 5.1. The posted cost drops to $0.042 per million input tokens, compared to a range of $0.20 to $10 for standard models.
The demonstrations are deliberately unusual: piloting reactive bots in game situations taken from Doom, at a rate of ten requests per second for about seven dollars per hour, and selecting links in games of Wikiracing. Access is by gradually unlocking the existing waiting list.
Read the full article on the original source
Go to the source →