GPT-6 Astra Ultrafast arrived on October 1, 2026: that day, NVIDIA's official blog published "How NVIDIA GPUs Help Accelerate OpenAI's GPT-6 Astra Ultrafast," announcing that the new version running on NVIDIA Blackwell GPUs is available to developers through the OpenAI API starting today, with eligible ChatGPT Work and Codex users getting access at the same time. Compared with standard Astra, Ultrafast generates tokens up to 8 times faster — coding loops spin faster, tool calls chain together more smoothly, and interactive apps respond noticeably quicker.
What Ultrafast is
Ultrafast is not an entirely new model, but a high-speed variant of GPT-6 Astra, built for Blackwell. No new architecture to learn, no new interface — the same Astra, just much faster generation.
Where the 8x comes from
The 8x figure doesn't come from piling on hardware; it is the product of OpenAI's inference optimization combined with the Blackwell architecture's capabilities. Philippe Tillet, OpenAI's head of inference, notes that NVIDIA's deep investment in tooling and documentation has made OpenAI's models very good at writing high-performance kernels for Blackwell and future Rubin GPUs — and Astra Ultrafast turns that acceleration into tangibly faster responses when agents write code and call tools. Uday Ruddarraju, OpenAI's Compute CTO, adds that the team keeps optimizing inference software on NVIDIA GPUs with its own in-house models, and the platform's programmability is the key to this speedup.
Who benefits first
The speed gains land first in three kinds of scenarios: the coding agent's "write code — run tests — fix bugs" loop, where every step waits on generation, so 8x speed makes the whole loop turn faster; tool-call-heavy tasks, where less waiting between calls keeps the agent's "think — act — think again" rhythm flowing; and latency-sensitive apps like chat and code completion, which feel noticeably more responsive.
What it means for developers
Notably, this speedup goes beyond a one-time launch boost: OpenAI keeps refining its inference software on NVIDIA GPUs with its own models, while the programmable platform lets the same infrastructure be reused across training, inference, and reinforcement learning, raising utilization — so Ultrafast's speed advantage will likely keep growing as the software evolves. Developers who want to try it can check the Ultrafast guide in OpenAI's developer documentation, which covers access, pricing, and implementation details.
