The Pulse
NVIDIA Puts OpenAI’s Astra Ultrafast on Blackwell GPUs
NVIDIA says OpenAI’s GPT-6 Astra Ultrafast runs on Blackwell GPUs and can generate tokens up to eight times faster than Astra Standard mode. The service is available through the API and to eligible ChatGPT Work and Codex users, with API acc
AI.info Team · Written with AI ·
Blackwell runs Astra’s faster mode
On October 1, NVIDIA said OpenAI’s GPT-6 Astra Ultrafast runs on NVIDIA Blackwell GPUs, bringing a faster service tier to the model through the OpenAI API and eligible ChatGPT Work and Codex accounts. NVIDIA claims Ultrafast generates tokens up to eight times faster than Astra’s Standard mode. The announcement ties that increase to inference software optimizations developed by OpenAI for NVIDIA’s GPU architecture.
The eightfold figure is a claim about token generation, not a promise that every task finishes eight times faster. NVIDIA’s post does not identify the workloads or measurement method behind the comparison. For coding agents that repeatedly generate text, call tools and wait for the next response, faster token output could reduce pauses between steps, but it does not by itself establish an equivalent improvement in total task time.
OpenAI says its models helped tune inference
NVIDIA describes the work as more than a hardware deployment: OpenAI used its own models to help refine inference software running on Blackwell GPUs. The post says the programmable platform let teams test and implement those changes, with the aim of making model responses faster and infrastructure more productive. It does not specify which Blackwell GPU systems OpenAI uses, how many GPUs are involved or where they are deployed.
“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. “Astra can turn that knowledge into high-performance kernels that make NVIDIA hardware compelling across the full frontier of latency, throughput and cost. With Astra Ultrafast, that means faster model responses as agents write code, use tools and work through complex tasks.”
Philippe Tillet, inference lead at OpenAI
Uday Ruddarraju, OpenAI’s chief technology officer of compute, described the same process in terms of model-assisted tuning. “We used our internal models to optimize inference on NVIDIA GPUs, and NVIDIA’s programmability helped us deliver the acceleration behind Astra Ultrafast,” he said.
API access comes with limits and a higher rate
OpenAI’s API documentation says Ultrafast is available for GPT-6 Astra to API users, subject to low rate limits. Developers select the ultrafast service tier in a Responses API request. OpenAI recommends a persistent WebSocket connection for agentic applications with frequent tool calls, because network overhead can eat into latency gains; HTTP requests are also supported.
The faster tier costs more than Standard mode, OpenAI says, though its guide directs customers to a separate pricing table for current token rates. The guide lists default Ultrafast limits of 500,000 tokens per minute for API usage tiers 1–3, 1 million for tier 4 and 5 million for tier 5. Those are throughput limits, not a measure of how quickly a single response arrives.
Regional processing is restricted
OpenAI’s documentation says Ultrafast supports global processing and U.S. data residency, but not EU or other non-U.S. regional processing endpoints. That restriction, the higher service cost and the API rate limits will shape which applications can use the mode, even where faster output would help.
NVIDIA’s announcement presents the release as a joint software-and-hardware optimization, but offers no independent benchmark results or deployment scale. The concrete change is narrower: developers can request GPT-6 Astra Ultrafast through the API, while eligible ChatGPT Work and Codex users can select it in their products.