OpenAI GPT-6 Astra Ultrafast Runs Up to 8x Faster on NVIDIA Blackwell

Summary

OpenAI released GPT-6 Astra Ultrafast on October 1, 2026, through its API and for eligible ChatGPT Work and Codex users. Model based inference optimizations exploiting NVIDIA Blackwell GPUs deliver up to 8x faster token generation than Astra Standard. The gain can shorten coding agents’ edit, test and debug cycles, reduce delays between tool calls, accelerate workflows that write code, use tools, check results and select next steps, and make interactive applications more responsive.

OpenAI inference lead Philippe Tillet said NVIDIA tooling and documentation helped its models program Blackwell and Rubin GPUs, enabling Astra to produce high performance kernels spanning latency, throughput and cost. Compute CTO Uday Ruddarraju said OpenAI used internal models to optimize NVIDIA GPU inference and continues testing and implementing improvements after deployment. NVIDIA’s programmable infrastructure can support training, inference and reinforcement learning, letting teams shift compute as demand changes, improve utilization and avoid workload specific overprovisioning. OpenAI’s Ultrafast guide contains access, pricing and implementation details.