OpenAI Opens GPT-6 Astra to Professional and Enterprise Users
OpenAI has announced that its flagship model GPT-6 Astra is now available for paid seats of ChatGPT Pro, Business, and Enterprise, and is also being integrated into API, Microsoft Azure, and Amazon Bedrock. The company stated that usage will count against existing subscription limits, with additional usage available for purchase.
Astra is positioned as a next-generation model for computer operations, browsing, software engineering, cybersecurity, science, and professional work. Official performance metrics include approximately 98% on FrontierMath tier 4, about 99.9% on ARC-AGI-3, and 100% on ExploitBench, with a context window of around 1.05 million tokens and knowledge cutoff around April 30, 2026. The API is priced at about $10 per million inputs and $50 per million outputs, with additional charges for long contexts.
Pro, Business, and Enterprise users will also receive GPT-6 Astra Pro; enterprise workspaces are disabled by default and must be enabled by an administrator. Business Premium seats are approximately $100 per month, offering higher limits and no five-hour cap. Free and entry-level tiers will continue to use lower-cost models like GPT-5.6 Luna. Codex must be upgraded to a specified CLI version to be called.
Before the full rollout, trusted companies like Daybreak will receive access. The sixth generation was launched about two months after the previous GPT-5.6, with delays due to model attacks on Hugging Face and other incidents leading to increased security measures. The company stated that its cybersecurity capabilities are internally marked as "Critical," and the external version will tighten certain usage requests.
In market dynamics, buyers aim to transition from chat-based interactions to high-priced, operating system-level workflows, while sellers charge based on tasks rather than per sentence in a closed-source lab. This represents a computational resource redistribution driven by product upgrades: funds are shifting from mid-tier reasoning in the 5.6 series to flagship tiers capable of operating computers and handling long tasks. Beneficiaries include OpenAI and cloud distribution partners who can sell seat prices and credits, while small to medium developers face pressure as they must complete equivalent work with more expensive output tokens and are often locked out by default settings.
Source: Public Information
ABAB AI Insight
Astra is not just about making chat smarter again; it connects the model to keyboards, mice, browsers, and code repositories. OpenAI includes both "most aligned" and "capable of hacking systems" in the same release note, indicating that product managers are focused on enterprise work orders while security teams prioritize refusal rates. The default closure for enterprises means that new capabilities are initially set as administrator switches rather than user rights.
The capital path involves overselling subscription limits with credits. The $100 and $200 tiers will access the model first, followed by the $20 Plus later, while the API will generate high margins by pricing long documents at $10/$50 for input/output. Azure and Bedrock sell the same weights into existing cloud contracts, reducing procurement friction for enterprises. Money flows from the number of seats in IT budgets to agents capable of completing tax filings, writing code, and interacting with screens.
This is not a textual leap from GPT-4 to GPT-5 but rather a transition from document processing to Power Automate in Office: once capabilities can operate computers, the pricing unit shifts from words to tasks. Anthropic and Google remain in the same agent race, but OpenAI locks distribution rights into Microsoft and Amazon invoices through cloud distribution. The industry position shifts from "Q&A interface" to "default-closed digital employees."
Structural changes belong to technological substitution. What is being replaced are basic computer operations and long-form documentation, not all knowledge work. The mechanism is that once evaluation scores elevate the model into critical tasks, the real gatekeepers become seats, administrator switches, and cybersecurity refusals; those who can pass these three gates will charge by task, while those who cannot will continue to pay mid-tier bills for 5.6.
ABAB News · Cognitive Laws
Models that can operate computers are priced by task rather than by word.
Default closure for enterprises is the true product launch.
The strongest capabilities often come with the strictest switches.