OpenAI launches Ultrafast for GPT-6 Astra as NVIDIA announces Blackwell deployment
OpenAI is introducing a paid, low-latency Ultrafast mode for the GPT-6 Astra API. The company claims speeds of up to eight times faster than Standard, but lower limits and data-processing restrictions apply at launch.
GPT-6 Astra Ultrafast is a new service mode in OpenAI’s API focused on low response latency. According to the documentation, it is generally available for the GPT-6 Astra model, although lower limits apply at launch. OpenAI states that the mode is “up to 8×” faster than Standard. However, this is the maximum stated benefit, not a guaranteed speedup for every request.
The mode is intended for paid use, and OpenAI also warns that it is more expensive than standard processing. Specific pricing and production-use costs will therefore be especially important for developers operating a high volume of API calls.
GPT-6 Astra Ultrafast has limited access in ChatGPT
At launch, Ultrafast is available in ChatGPT Work and Codex to users on the Pro $500 plan and eligible Enterprise and Edu workspaces. Plus, Pro $100, Pro $200, and Business subscribers do not have access.
The access breakdown shows that OpenAI is offering the low-latency mode separately from the model itself and, in the initial phase, is directing it primarily to the most expensive individual plan and selected organizations. The mode is generally available for the API, but with lower limits at launch.
Maximum eightfold acceleration is not a universal result
The “up to 8×” figure compares Ultrafast with the Standard service mode. Actual response time in a specific application can be affected by the network connection, output streaming, tool use, the length of the generated response, and the limits of the account.
The verified materials did not include independent benchmarks confirming this maximum figure in typical applications. When evaluating its benefit for production deployment, it will therefore be important to compare latency and throughput on specific workflows rather than relying only on the stated maximum.
Lower latency is particularly relevant for interactive systems, coding agents, and agentic workflows involving repeated tool calls. In such tasks, individual waits for the model can add up to the total task-completion time.
API does not support regional endpoints in the EU
OpenAI also states a data-processing restriction in its documentation. Ultrafast for the API supports only US data residency and global processing. The mode does not support regional endpoints in the European Union.
For organizations that need to use specific regional endpoints, this is a practical limitation when deciding whether to deploy the new mode. The documentation also identifies Ultrafast as a more expensive option, so its cost will also need to be considered before production use.
NVIDIA cites Blackwell GPUs
On October 1, 2026, NVIDIA announced that GPT-6 Astra Ultrafast runs on NVIDIA Blackwell GPUs. According to the company, performance is also improved by inference optimizations from OpenAI.
This claim about specific hardware deployment comes from NVIDIA. OpenAI does not separately specify the infrastructure in its verified documentation. Further developments will show the mode’s pricing, whether access expands to additional plans and workspaces, independent measurements against Standard, and any expansion of regional processing beyond the United States.
Sources
- OpenAI API Documentation — Ultrafast mode – Confirms Ultrafast availability for GPT-6 Astra, the claim of speeds up to 8× faster, API limits, and the regional-processing restriction.
- OpenAI Help Center — ChatGPT Work and Codex – Confirms which plans and workspaces have access to Ultrafast in Work and Codex at launch.
- NVIDIA Blog — How NVIDIA GPUs Help Accelerate OpenAI’s GPT-6 Astra Ultrafast – Confirms NVIDIA’s October 1, 2026 announcement that the mode runs on Blackwell GPUs and uses inference optimizations.
Verified and updated: 10/02/2026 06:21
