The rumors that have circulated through the halls of Silicon Valley for months have finally been confirmed: OpenAI has officially released GPT-6 Astra. The company is positioning this new frontier model as a watershed moment in artificial intelligence, one that it believes marks the definitive onset of artificial generalized intelligence (AGI). For OpenAI, this represents the culmination of its long-sought objective: to build highly autonomous systems capable of outperforming humans across the majority of economically valuable tasks.
During a closed-door press briefing held earlier today, OpenAI co-founder and president Greg Brockman provided a remarkably direct assessment of the model’s arrival. Ending the session with a declaration that resonated through the room, he stated simply, "Welcome to the AGI era."
This framing is unusually bold, even by the high-stakes standards of frontier AI development. While the label of AGI will undoubtedly spark intense academic and public debate, the practical implications for enterprises are far more immediate and concrete. OpenAI is marketing GPT-6 Astra as the foundation for a new era of computing—one where the traditional necessity of clicking through menus and typing on a keyboard may become entirely optional. In internal materials shared with the media, the company explicitly characterizes Astra as "the world’s best computer use model."
Unlike previous generations of AI, which required developers to build dedicated API integrations for every specific application, Astra is engineered to navigate software in the same way a human would. It can move fluidly across web browsers, spreadsheets, specialized desktop applications, and collaborative suites. The model is capable of producing finished documents and presentations, but more importantly, it can execute complex, multistep workflows from start to finish, rather than merely offering guidance on how a user might complete them.
To illustrate this shift, the company released a promotional video that juxtaposed the history of computing with its current leap. It began with a grainy, nostalgic demo from the 1980s showing a person asking a computer to draw a simple yellow circle. The video then cut to the present day, featuring OpenAI employees interacting with Astra through voice alone. They commanded the AI to transform that yellow circle into a complex rocket ship, evolve it into a full 3D game, and even list items for sale on eBay—all without the human ever touching a mouse.
OpenAI confirmed that Astra is beginning its rollout this Thursday to enterprise customers through its gated access program, Daybreak. Availability will expand in the coming days to ChatGPT Plus, Pro, Business, and Enterprise tiers, as well as through the OpenAI API and major cloud platforms like AWS Bedrock and Microsoft Azure.
From Answering Questions to Operating Computers
The enterprise value proposition for Astra centers on its ability to act as a digital operator. OpenAI reports that the model is adept at filling out online forms, updating customer relationship management (CRM) records, managing calendars, conducting deep web research, and drafting professional communications. Furthermore, it can manipulate complex spreadsheets, analyze scientific data in Python notebooks, work within business intelligence tools like Power BI, and troubleshoot or install software.
This capability signals a potential transformation in enterprise AI architecture. For the duration of the generative AI boom, corporations have been burdened with the technical debt of connecting models to corporate systems via APIs, plugins, and retrieval systems. Brockman argued that computer-use agents like Astra could bypass much of this friction because modern software is already designed for a general-purpose intelligence: the human user.
"We’ve been bottlenecked over this gigantic era by people writing connectors and very painstakingly building these connections into all these tools that people can already use," Brockman explained during the briefing. He suggested that with a sufficiently capable agent, users can skip the middleman, allowing the model to "zip through spreadsheets, fill out forms, and navigate across web pages" as if it were an employee sitting at a terminal.
This vision traces back to the earliest days of OpenAI, where researchers first discussed the concept of training an agent using the same fundamental inputs and outputs available to humans: pixels, keyboards, and mice. According to Brockman, the company has finally achieved the first iteration of an agent that performs these tasks with genuine, high-utility results. The performance metrics are telling; on an offline subset of OSWorld 2.0, Astra scored 72.6% while completing tasks in approximately 40 minutes, a significant improvement over its predecessor, GPT-5.6 Sol, which scored 65.7% with an average task time of 75 minutes.
This shift—moving from prompting an AI to supervising an AI—is what OpenAI believes will matter most to businesses. As researcher Mia Glaese noted during the briefing, the goal is to enable people to delegate increasingly complex work across various applications, allowing humans to direct the work at a higher, more strategic level.
A Massive Leap in Training Scale
Aidan Clark, an OpenAI researcher, described the development of Astra as the company’s most ambitious training run to date. Astra is the first model pretrained using more than 100,000 DBUs at the company’s Stargate infrastructure. Furthermore, it marks the first time that previous iterations of OpenAI’s models have played a supervisory role in the training of their successor.
The results are striking across various benchmarks. Astra achieved a 97.6% score on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1, and 100% on ExploitBench. Perhaps most notably, it reported a 98.6% score on ARC-AGI-3, a benchmark designed to test whether an AI can generalize to unfamiliar problems rather than simply regurgitating training data.
However, these figures have ignited a debate within the AI community regarding the nature of intelligence measurement. The 98.6% score on ARC-AGI-3 highlights a growing friction in the industry: how do we define the "object" being measured? Is it the raw foundation model, the model bolstered by persistent memory, or the entire deployed agent system?
NVIDIA recently demonstrated that its own "Agentic Variation Operators" (AVO) architecture could achieve a 100% score on ARC-AGI-3 using a base model that would otherwise score around 30%. By adding persistent memory, tool-use, and feedback loops, NVIDIA proved that long-horizon capabilities often emerge from the surrounding system architecture rather than the model weights alone. This has led some critics to argue that such benchmarks are becoming less honest as they are increasingly "gamed" by elaborate harnesses.
OpenAI acknowledges the complexity of this debate. While they are confident in Astra’s capabilities, they admit that the definition of AGI remains "fuzzy." For Brockman, the evidence is in the utility: if a model can perform across a wide spectrum of difficult, real-world tasks, the distinction between foundation model and system architecture may matter less to the enterprise than the final, audit-worthy result.
Governance and the Challenge of Autonomy
The very capabilities that make Astra a powerful tool for enterprise productivity also introduce a profound governance challenge. A standard chatbot generates text for human inspection, but an agent capable of operating a computer can actively modify records, send information, and manipulate files.
"Even as models can do more things autonomously, we have to be able to trust them more," Glaese said. OpenAI has adopted a defense-in-depth approach to safety, moving away from simple "refusal layers" toward a comprehensive framework that includes classifiers, security controls, and real-time monitoring.
The company revealed that it paused some frontier training for approximately two weeks following an external security incident, not because Astra itself was compromised, but to ensure that their internal safety and monitoring infrastructure could keep pace with the model’s rapidly evolving capabilities. For enterprise leaders, this underscores a shift in risk management: governance is no longer just about the prompt, but about the sequences of actions, the authorization boundaries, and the ability to detect suspicious behavior before it results in a breach.
OpenAI has specifically designated Astra as its first model to reach the "Critical" cybersecurity threshold under its Preparedness Framework. This means the model is capable of identifying previously unknown vulnerabilities and developing exploit chains without human guidance. To manage this "dual-use" risk, OpenAI is initially restricting the most advanced cyber capabilities to trusted defenders and organizations responsible for protecting critical infrastructure.
Ultimately, OpenAI’s message is that the transition to the AGI era will likely be an economic one, measured by the delegation of work rather than a single triumphant test score. If Astra can consistently and reliably perform professional tasks while remaining within defined safety boundaries, organizations will naturally restructure their workflows to accommodate it.
As Brockman suggested, the true test of this era will be found in the coming year. If enterprises begin to hand over significant, consequential work to these systems, the point at which we entered the AGI era may become obvious only in the rearview mirror. For now, OpenAI is focused on the immediate task: proving that Astra can provide the reliability and utility that businesses demand, regardless of the label they choose to apply to it.
