GPT-6 Astra Ultrafast arrives on Nvidia Blackwell with up to 8x faster output

Nvidia says Astra Ultrafast is available in the API and to eligible ChatGPT Work and Codex users. Its 8x claim measures token generation against Standard.

Oct 2, 2026

GPT-6 Astra Ultrafast is available now in the OpenAI API and to eligible ChatGPT Work and Codex users, according to Nvidia’s October 1 announcement. Nvidia says it runs on Blackwell GPUs and offers up to 8x faster token generation than Astra Standard mode.

For builders, Nvidia points to shorter waits as coding agents write edits and generate responses between tool calls. Those savings can recur across an edit, test and debug cycle.

OpenAI’s inference work uses its own models

Nvidia attributes the acceleration to inference software optimizations that use Blackwell’s capabilities. OpenAI compute chief technology officer Uday Ruddarraju says the company used internal models to optimize inference on Nvidia GPUs. The post describes continued work on that software after deployment.

The announcement leaves ChatGPT Work and Codex eligibility at “eligible users” and directs developers to an Ultrafast guide for pricing and implementation details.

Claim check

Measure the whole task before expecting an 8x gain

The lead image, Nvidia’s published rocket demo comparison, labels Standard at 17.0 seconds and Ultrafast at 6.4 seconds. Dividing those displayed times gives about 2.7. That comparison measures the ratio of the displayed durations. The headline claim measures token generation, so the two numbers answer different questions.

The image alone does not establish output token counts or matching work, and the announcement gives no benchmark setup for the 8x maximum. Treat the displayed demo as an example rather than a forecast for your agent.

To evaluate a switch from Standard, run the same coding tasks in both modes. Record generation time separately from tool execution and test runs, then compare total time, successful task completion and cost per completed task. Faster generation has the most room to help when it accounts for much of the wait. A slow test suite can still determine when the agent finishes.

Sources

NVIDIA Blog: How NVIDIA GPUs Help Accelerate OpenAI’s GPT-6 Astra Ultrafast. Source of the lead demo image, credited to Nvidia.