OpenAI’s GPT-6 Astra Ultrafast is available through the OpenAI API and to eligible users of ChatGPT Work and Codex, according to NVIDIA. The model runs on NVIDIA Blackwell GPUs. NVIDIA says Ultrafast can generate tokens up to eight times faster than GPT-6 Astra in Standard mode. The announcement does not give the conditions or measurements behind that comparison, so the figure should be treated as a company-reported maximum, not a guaranteed speedup for every request.
The main intended use is work that involves repeated responses and tool calls. For example, a coding agent can write code, run a tool, inspect the result and decide what to do next. NVIDIA says faster generation can shorten those waits and make interactive applications feel more responsive. It does not say how much time users will save in a particular task.
How is the speed achieved?
OpenAI says it used its own models to help optimize the inference software that runs on NVIDIA GPUs. Inference is the process of generating a response from a trained model. According to NVIDIA, the work takes advantage of Blackwell’s programmability to develop and implement improvements to that software. OpenAI inference lead Philippe Tillet said the company used NVIDIA’s tooling and documentation to make its models better at programming Blackwell and Rubin GPUs, and that Astra can use this knowledge to produce high-performance kernels—software that runs operations on the hardware.
The announcement describes this as ongoing work: OpenAI says it can continue refining the software after deployment. NVIDIA also argues that a programmable platform can let teams reuse infrastructure for training, inference and reinforcement learning as workloads change. The blog does not provide independent test results, detailed benchmark settings or a breakdown of the hardware used for Ultrafast.
Who can use it?
Developers can access GPT-6 Astra Ultrafast through the OpenAI API. NVIDIA says eligible ChatGPT Work and Codex users can also use it, but does not specify eligibility rules in the post. It points developers to an Ultrafast guide for pricing and implementation details; the blog itself gives no price or account of how costs compare with Astra Standard. That leaves an important practical question open: faster generation may help in time-sensitive workflows, but users will need the guide and their own usage details to assess cost and access.
Why it matters
For developers using coding agents or other tool-based workflows, faster responses could reduce pauses between steps, though the post does not quantify time saved in real tasks. API pricing and the conditions behind the up-to-8x figure are not included, so users cannot judge cost or expect the maximum speedup from this announcement alone.