OpenAI has launched GPT-6 Astra Ultrafast, powered by NVIDIA Blackwell GPUs, making the model available via API and to eligible ChatGPT Work and Codex users. According to OpenAI, inference optimizations tailored to NVIDIA’s Blackwell architecture enable Ultrafast to deliver up to 8x faster token generation compared to the standard Astra mode.
The performance boost is intended to shorten developer edit-test-debug loops, decrease tool-call latency, and enable smoother interactive agentic workflows. OpenAI inference lead Philippe Tillet stated that internal models were used to optimize code for programming Blackwell and Rubin GPUs, generating high-performance kernels that enhance latency, throughput, and hardware cost efficiency.
NVIDIA’s programmable platform allows teams to reuse compute infrastructure across training, inference, and reinforcement learning workloads. According to OpenAI CTO of Compute Uday Ruddarraju, the partnership leveraged internal AI models to refine GPU inference software, delivering continuous hardware utilization improvements.
Why it matters
Developer agents benefit from 8x faster token generation, significantly shortening edit-test-debug latency loops in coding tasks.
Co-optimization between frontier model software and NVIDIA Blackwell GPUs underscores hardware-software integration as a critical performance lever.
Source: blogs.nvidia.com



