GPT-6 Astra Ultrafast Claims Up to 8x Faster Token Generation on NVIDIA Blackwell GPUs

OpenAI's GPT-6 Astra Ultrafast is available now in the OpenAI API and to eligible ChatGPT Work and Codex users, running on NVIDIA Blackwell GPUs. According to NVIDIA, Ultrafast delivers up to 8x faster token generation than the Astra Standard mode.

What was announced

The details below come from NVIDIA's blog post, not from an independent test.

• Availability: Live now in the OpenAI API and for eligible ChatGPT Work and Codex users. The source does not define "eligible."
• Hardware: NVIDIA Blackwell GPUs.
• Speed: Up to 8x faster token generation than Astra Standard mode. This is a vendor claim, and "up to" means the figure is a ceiling, not a typical result.
• Method: NVIDIA attributes the gain to inference optimizations in OpenAI's models that use the capabilities of the Blackwell architecture.

Ultrafast is framed as a mode, not a new model

NVIDIA describes Ultrafast as a mode relative to "Astra Standard." It does not describe a separate trained model with its own weights. Based on the source, this reads as a serving configuration of GPT-6 Astra optimized for latency. We have not seen a distinct model identifier or a separate model card.

What is not disclosed

The source excerpt leaves out several figures developers will need:

• Pricing per 1M tokens: not yet disclosed
• Context window: not yet disclosed
• Benchmark scores: not yet disclosed, including whether Ultrafast matches Standard on quality
• Parameter count and training cutoff: not disclosed
• Baseline throughput: the 8x figure has no stated tokens-per-second baseline, batch size, or prompt length

The source text was also truncated at the section on developer benefits, so further details may exist in the full post.

What this means

The headline number matters less than the missing conditions. An 8x gain over Standard is large, but without a baseline throughput, latency-to-first-token data, or quality comparison, builders cannot yet judge it. The key open questions are whether Ultrafast costs more per token than Standard and whether output quality is unchanged. Faster serving often comes with a price premium or a tradeoff, and neither has been confirmed here.

The release also shows how closely frontier model providers and hardware vendors are now aligned on inference. OpenAI names Blackwell as the platform for its fastest tier, and NVIDIA is publicizing the result. That points to inference speed becoming a competitive axis alongside model quality, particularly for agentic coding tools like Codex, where many sequential generation steps make latency the bottleneck.

Until OpenAI publishes its own documentation with pricing, limits, and measured performance, treat the 8x figure as a ceiling claimed by NVIDIA and test it against your own workloads.