NVIDIA Speeds Up GPT-6 Astra Ultrafast on Blackwell

A dedicated inference tier uses AI models to write GPU kernels on Blackwell chips, though total response time drops 2.6x at $300 per million tokens.

NVIDIA and OpenAI have published the technical details behind GPT-6 Astra Ultrafast, a specialized inference tier designed for autonomous agent workloads. Announced on October 1, 2026, the mode runs exclusively on NVIDIA Blackwell graphics processors and delivers an output token rate up to eight times faster than standard execution. The primary technical twist is how the software was assembled: OpenAI used its own artificial intelligence models to write and optimize low-level compute kernels directly for NVIDIA silicon.

AI models generating code for silicon

According to Philippe Tillet, head of inference at OpenAI, the lab ran internal models over NVIDIA technical documentation and development tools to produce custom kernels. These low-level routines direct math operations across GPU cores and ended up outperforming manually written code from human engineers.

Uday Ruddarraju, OpenAI head of compute technology, pointed to the programming flexibility of Blackwell hardware as the key factor behind the automated tuning. While engineering teams conducted exploratory tests with NVIDIA’s upcoming Rubin architecture, commercial support at launch remains confined to Blackwell systems.

How GPT-6 Astra Ultrafast performs in practice

The 8x multiplier emphasized by NVIDIA measures output token throughput: the rate at which text or code appears once generation starts. It does not translate into queries finishing eight times faster from start to finish, because overall response time still includes prompt ingestion and initial processing latency.

Benchmark figures published by NVIDIA show that end-to-end task duration fell by roughly 2.6x under comparable test workloads, dropping from 17.0 seconds in standard mode to 6.4 seconds in Ultrafast. For autonomous coding tasks in Codex, that burst speed still offers concrete benefits. Agents repeatedly cycle between writing script blocks, executing commands in terminal sessions, and parsing error logs, meaning tighter token delivery cuts down cumulative idle time.

Steep pricing and data residency limits

That speed boost carries a hefty price tag. In OpenAI’s developer documentation, calling the tier via the service_tier: ultrafast flag on gpt-6-astra (with OpenAI advising WebSockets for agent connections) costs $300 per million output tokens. That represents a six-to-eightfold premium over base pricing. Within ChatGPT, access is restricted to $500 monthly ChatGPT Pro subscribers, alongside enterprise and educational contracts.

Beyond financial costs, deployment comes with geographical restrictions. OpenAI documentation confirms that Ultrafast relies on global routing and mandates primary data storage within the United States. Organizations bound by strict local data governance rules in the European Union cannot adopt the service at launch, and OpenAI has not shared a target date for EU-hosted infrastructure.