OpenAI's GPT-6 Astra Ultrafast runs 8 times faster on NVIDIA GPUs
The new mode speeds up coding agents and interactive apps by reducing token generation time.
OpenAI has released GPT-6 Astra Ultrafast, a faster version of its language model. It is now available in the OpenAI API and for specific ChatGPT Work and Codex users. The system runs on NVIDIA Blackwell GPUs.
The main change is speed. Ultrafast generates tokens up to 8 times faster than the standard Astra mode. This matters most for developers using coding agents. When an AI writes code, tests it, and fixes errors, every second counts.
NVIDIA’s deep investment in tooling and documentation has enabled us to make our models exceptionally good at programming Blackwell and Rubin GPUs.
Imagine an agent that checks its work and decides the next step. Faster responses make these loops feel more responsive. It reduces the waiting time between tool calls. The goal is to make interactive applications feel snappier.
"NVIDIA’s deep investment in tooling and documentation has enabled us to make our models exceptionally good at programming Blackwell and Rubin GPUs," said Philippe Tillet, inference lead at OpenAI.
OpenAI uses its own models to refine the software running on these chips. This process helps optimize how the hardware handles data. It allows the team to improve performance over time without changing the physical equipment.
The NVIDIA platform is programmable. This means developers can reuse the same infrastructure for training, inference, and reinforcement learning. It helps avoid buying extra hardware for every new task.
Developers can access the new mode through the API today. OpenAI has published a guide with pricing and implementation details. The speed boost is designed for time-sensitive workflows.
This release focuses on inference speed for existing model capabilities. The source does not mention new reasoning abilities or changes to the model's core knowledge. It is an optimization of how the model runs, not a new brain.
Source: NVIDIA Blog