OpenAI released GPT-6 Astra on September 3, 2026, first to a limited group of trusted partners, then to paid users a day later. The company is calling it their most powerful model to date, and the feature list gives that claim some substance: state-of-the-art performance across software engineering, science, professional work, and — most notably — a “computer use” capability that lets the model navigate a computer the way a human would.

Access is rolling out through enterprise customers in OpenAI’s Daybreak programme first, ahead of broader ChatGPT and API availability.

What GPT-6 Astra Actually Does

The headline feature is computer use. Rather than accepting inputs and generating text responses, Astra can operate software directly — clicking, typing, reading screens, filling forms, and executing multi-step tasks in applications. This puts it in the same territory as Anthropic’s Computer Use and similar agentic capabilities that have been maturing across the industry.

This matters because it changes the category of task AI can take on. Text generation, code completion, and analysis are already embedded in most AI tools. Computer use means AI can operate the actual systems businesses run on — without APIs, custom integrations, or specialized tooling. An agent that can use software the way a human uses it is an agent that can take on workflows that are currently human-only.

Beyond computer use, Astra leads on benchmarks across software engineering, scientific reasoning, and professional work. It is multimodal, has a context window exceeding one million tokens, and supports 128K token output — meaning it can work with very large documents and produce long, detailed outputs without truncation.

Pricing and Access

OpenAI has priced GPT-6 Astra at $10 per million input tokens and $50 per million output tokens. This aligns it with Anthropic’s Fable 5.1 (which also became generally available this month at the same price point), and positions both as premium enterprise options rather than everyday consumer tools.

Enterprise customers in the Daybreak programme get priority access. Broader API access for developers and rollout to paid ChatGPT plans follows. OpenAI has restricted certain capabilities in the initial release — particularly around cybersecurity prompts — citing the model’s advanced capabilities in that domain.

The Safety Context

GPT-6 Astra did not ship on schedule. Following a Hugging Face-related incident in July 2026, OpenAI delayed the release to build in additional safeguards. The restricted cybersecurity access is part of that response.

This is relevant for enterprise buyers evaluating the model. OpenAI is clearly aware that a model with this level of capability — especially computer use and advanced cybersecurity knowledge — requires more careful access controls than previous generations. Organizations deploying it will need to think about what it can access, what it can do autonomously, and what oversight mechanisms they put in place.

What This Means for Business

GPT-6 Astra raises the bar on what enterprise AI can do. Here is how to think about it in practical terms.

Computer use changes automation scope. If your organization has workflows that require a human to sit at a computer and operate software — because that software has no API, or because the task is too complex to script — those workflows are now candidates for AI automation in a way they were not before. The constraint was never that people lacked willingness to automate; it was that the tools could not do it. That constraint is shrinking.

The premium pricing reflects premium use cases. At $10/$50 per million tokens, Astra is not a tool for high-volume, low-complexity tasks. It is priced for work where the value per output is high — strategic analysis, complex software projects, scientific research, professional services work where quality matters more than cost per query.

Benchmark leadership has a short shelf life. Anthropic’s Fable 5.1, Google’s models, and others are competing at the same tier. No single model holds the top position for long. What matters more than which model is “best right now” is having the internal capability to evaluate, switch, and deploy models as the landscape shifts. Organizations that have built AI-literate teams and flexible infrastructure will capture value from each new release; those that are still running point-in-time pilots will always be catching up.

Computer use needs governance. An AI that can operate your computer systems needs guardrails. What accounts can it access? What actions can it reverse? What approval workflows are in place? These are not theoretical questions — they are prerequisites for safe deployment.

For business leaders working through what the current generation of AI means for their operations, Enterprise DNA’s Omni advisory service helps map capability to context — what’s worth using now, how to build internal competence, and how to govern AI deployments at scale. Book a discovery call with Sam McKay to work through what GPT-6 and its competitors mean for your organization.