NVIDIA reports that OpenAI’s GPT-6 Astra Ultrafast is now available through the OpenAI API and to eligible ChatGPT Work and Codex users. The mode runs on NVIDIA Blackwell GPUs and uses inference optimizations in OpenAI’s models that draw on the Blackwell architecture. According to the description, Ultrafast delivers up to 8x faster token generation compared with the Astra Standard mode.
Why it matters
Faster token generation can reduce response latency for applications built on the model, which is relevant for interactive and high-throughput use cases. The stated performance gain is tied to hardware-level inference optimizations on the Blackwell architecture.
Who should care
Developers using the OpenAI API, along with eligible ChatGPT Work and Codex users, are the audiences named as having access to the Ultrafast mode.