OpenAI’s September 2026 changelog records two separate developments: GPT-6 Astra launched on September 3, while its Ultrafast API service tier arrived on September 29. Astra is a model name here. Google’s Project Astra is a different initiative, described by Google DeepMind as a research prototype; the two should not be confused.
OpenAI’s model documentation lists a context window of 1,050,000 tokens and a maximum response of 128,000 tokens. It accepts text and images and returns text. Native audio and video are not supported by this model. A large context window means more material can fit into a request; it does not prove that every detail will be remembered or interpreted correctly.
The GPT-6 developer guide describes workflows that can call tools asynchronously and accept new instructions while work is running. In practical terms, a developer could let an agent continue useful work while another tool result is pending, then redirect its priorities. The application must still provide the tools and permissions; a model announcement does not grant access to someone’s files or accounts.
Standard API prices are US$10 per million input tokens and US$50 per million output tokens. Prompts above 272,000 input tokens raise input and cache rates twofold and output rates by half for the whole request. These are metered developer prices, not a ChatGPT subscription quote. Long documents, repeated attempts and tool fees can change the total bill.
OpenAI’s Ultrafast guide advertises up to eight times Standard-mode speed, with initially low API rate limits. It supports global processing and US data residency, but not EU or other non-US regional processing. The speed claim comes from OpenAI and should not be read as a guarantee for every job. Ultrafast is a higher-priced service tier rather than a new model.
For a practical evaluation, try the same bounded assignment on a smaller model and Astra: a code repair with tests, a sourced research brief, or a document with explicit acceptance criteria. Compare correctness, time and total cost. Keep human approval for consequential actions, inspect citations and run tests. This report reflects official documentation checked on October 3; the launches themselves occurred in September.