OpenAI releases GPT-6 Astra Ultrafast, a faster model running on Nvidia Blackwell

OpenAI has released GPT-6 Astra Ultrafast, a faster version of its Astra model built to run on Nvidia Blackwell, Nvidia’s newest GPU architecture. The model is available now in the OpenAI API and to eligible ChatGPT Work and Codex users, according to a post on Nvidia’s blog.

Nvidia says Ultrafast generates tokens, the chunks of text a language model outputs one at a time, up to eight times faster than Astra’s Standard mode. The gain comes from inference optimizations built into OpenAI’s models that are tuned to the Blackwell architecture, Nvidia says.

For software developers, faster output can shorten the edit-test-debug cycle that coding agents run through: writing code, calling a tool, checking the result, then deciding what to do next. Nvidia says that loop is where repeated delays add up, and that Ultrafast is meant to keep pace with it so agents and interactive applications feel more responsive.

Philippe Tillet, OpenAI’s inference lead, said Nvidia’s investment in tooling and documentation helped OpenAI’s models become skilled at programming Blackwell and Rubin GPUs. “Astra can turn that knowledge into high-performance kernels that make NVIDIA hardware compelling across the full frontier of latency, throughput and cost,” Tillet said, adding that with Ultrafast, that means faster responses “as agents write code, use tools and work through complex tasks.”

OpenAI also used its own models to help refine the inference software running on Nvidia GPUs, taking advantage of the chips’ programmability to test and apply improvements after the model was already deployed, Nvidia said. Uday Ruddarraju, OpenAI’s chief technology officer of compute, said the work with Nvidia is helping make AI “faster and more useful,” and that OpenAI used its internal models to optimize inference on Nvidia GPUs while Nvidia’s programmability helped deliver the acceleration behind Ultrafast.

Nvidia frames that programmability as a broader advantage beyond this one model: the same GPUs can be reused across training, inference and reinforcement learning as models change, which it says helps teams avoid buying more hardware than a given workload needs.

Developers can use GPT-6 Astra Ultrafast through the OpenAI API now. Nvidia’s post points to a separate Ultrafast guide for access, pricing and implementation details, which were not included in the material reviewed for this story. Not yet known is how Ultrafast performs across different coding and agent workloads in independent testing, and whether OpenAI will publish benchmark data beyond the eight-times figure it has given for token generation speed.