NVIDIA says GPT-6 Astra Ultrafast delivers up to eight times faster token generation than Astra Standard, using inference optimizations on Blackwell GPUs. Its October 1 announcement says the mode is available through the OpenAI API and to eligible ChatGPT Work and Codex users.

OpenAI’s documentation independently confirms that Astra Ultrafast is broadly available to API users, describing it as the API’s fastest service tier and advising developers to use it when speed justifies the higher cost. The eightfold figure is NVIDIA’s claim about token generation, not a guarantee that an entire coding task finishes eight times sooner.

Persistent connections matter

Developers select the tier with model: gpt-6-astra and service_tier: ultrafast. OpenAI supports both HTTP requests and WebSockets, but strongly recommends persistent WebSocket connections for agents making frequent tool calls. Without them, network overhead can reduce the latency gains.

That distinction matters for coding agents, whose workflows alternate between generating responses, executing tools and checking results. Faster generation can reduce one source of delay without eliminating tool execution or network time.

Access comes with constraints

The documented default limits are 500,000 tokens per minute for API usage tiers 1–3, one million for tier 4 and five million for tier 5. Ultrafast supports US data residency and global processing, but not EU or other non-US regional processing endpoints.

NVIDIA also says OpenAI used internal models to optimize inference software on its GPUs. For developers, the concrete change is an available latency-focused service tier, with cost, connection design and processing-region requirements still shaping where it fits.