NVIDIA says Blackwell GPUs are powering OpenAI’s faster “Astra Ultrafast” GPT-6 mode

NVIDIA on Oct. 1 detailed how it says its Blackwell GPUs are being used to accelerate OpenAI’s GPT-6 Astra Ultrafast, a faster-response mode now available through the OpenAI API and to eligible ChatGPT Work and Codex users. The company’s account focuses less on model training and more on inference, the compute that turns a model’s learned parameters into responses in production. NVIDIA says its platform and tooling helped OpenAI deliver a performance jump that is particularly important for applications where systems must repeatedly generate tokens, call external tools, and decide what to do next.

According to NVIDIA’s blog, Astra Ultrafast runs on NVIDIA Blackwell GPUs and is available now. NVIDIA said the Ultrafast mode can generate tokens up to 8x faster than Astra Standard. Tokens are the basic chunks of text (and code) that large language models produce step by step, so faster token generation translates into shorter wait times for users and for software agents that need to iterate quickly.

NVIDIA linked the speed improvement to a practical developer metric: reducing the time spent in edit-test-debug loops for coding agents. The company described workflows in which an agent writes code, uses a tool, checks the result, and then chooses a next action. In that kind of cycle, the delay between tool calls and the time to produce intermediate outputs can compound, making an interactive application feel slow even if the tools themselves respond quickly.

The blog also highlighted latency and throughput, the two core characteristics that affect how quickly and how consistently a deployed AI system can respond to requests. NVIDIA said OpenAI’s models tap into capabilities of the Blackwell architecture to generate “high-performance kernels,” software units that implement parts of the inference pipeline efficiently on the GPU. NVIDIA framed the result as delivering the acceleration needed to improve response times while balancing cost and operational efficiency.

A quote from Philippe Tillet, inference lead at OpenAI, attributed to the work NVIDIA helped enable: he said NVIDIA’s deep investment in tooling and documentation helped OpenAI make its models exceptionally good at programming Blackwell and Rubin GPUs. Tillet added that Astra can turn that knowledge into high-performance kernels that make NVIDIA hardware compelling across the “frontier of latency, throughput and cost,” and that Astra Ultrafast brings faster model responses for agent-based coding and complex tasks.

NVIDIA said the performance work continues after deployment as well. It described OpenAI using its own models to help refine inference software running on NVIDIA GPUs, taking advantage of the platform’s programmability to test and implement improvements over time. In this view, the overall system optimization is not treated as a one-time engineering project, but as an iterative cycle where inference libraries can be tuned as workloads and model behavior evolve.

A second quote came from Uday Ruddarraju, chief technology officer of compute at OpenAI. NVIDIA said Ruddarraju credited internal-model optimization of inference on NVIDIA GPUs and said NVIDIA’s programmability helped deliver the acceleration behind Astra Ultrafast. The message is that the speed gain is not only about having the right GPU hardware, but also about having software infrastructure that lets teams shape how inference executes at runtime.

Beyond the immediate Astra Ultrafast announcement, NVIDIA spent time explaining a broader architectural point: a programmable NVIDIA platform allows teams to reuse infrastructure across training, inference, and reinforcement learning as models change. NVIDIA said that flexibility can help improve utilization and avoid overprovisioning for each workload, an operational concern for data center operators that need to balance capacity and demand.

NVIDIA did not provide additional benchmarks beyond the claimed “up to 8x” token generation improvement, and it did not break out how much of the gain comes from model changes versus inference-engine optimizations. It also did not publish pricing for the mode in the blog post itself, pointing readers to an Ultrafast guide for access and pricing details. As with many performance claims in AI, the practical experience can depend on request patterns, batch sizes, deployment configuration, and how often applications trigger tool calls that shape an agent’s step-by-step flow.

For developers and operators, the main question to watch next is whether the Astra Ultrafast speed gains hold consistently across different application types and at scale. If the mode meaningfully reduces end-to-end latency in agent workflows, it could make tool-using coding assistants feel more interactive, and it may shift engineering priorities toward tighter inference loops. NVIDIA and OpenAI will likely be judged on how quickly they can translate this kind of acceleration into sustained capacity, reliability, and measurable user outcomes once the feature expands beyond initial eligibility groups.