The speculation has finally come to an end. OpenAI today officially released GPT-6 Astra, a groundbreaking frontier model that the company claims marks the long-awaited onset of Artificial General Intelligence (AGI). For years, OpenAI has pursued the ambitious goal of creating "highly autonomous systems that outperform humans at most economically valuable work," and with the debut of Astra, the organization believes it has finally cleared that horizon.
In a closed-door press briefing held earlier today, OpenAI co-founder and president Greg Brockman provided a remarkably direct assessment of the release. Addressing the significance of the moment, Brockman concluded the session by declaring, "Welcome to the AGI era."
This framing is unusually bold, even by the high-stakes standards of the frontier AI industry. Yet, for enterprise leaders and technology executives, the immediate value of Astra lies in something far more tangible than the philosophical debate over AGI: the transition from AI as a chatbot to AI as a computer operator. OpenAI is positioning GPT-6 Astra as the foundation for a new paradigm of computing, one where users and employees can move away from the manual drudgery of clicking through browsers, navigating spreadsheets, and typing on keyboards.
OpenAI’s internal materials describe Astra as "the world’s best computer use model." Unlike previous iterations that required developers to build custom API integrations for every specific application, Astra is engineered to navigate software much like a human does. It can operate across browsers, websites, and desktop applications, handling multistep workflows and producing finished documents or presentations, rather than simply offering instructions on how a human might complete them.
The company underscored this capability with a promotional video that bridged the gap between the past and the future. It opened with a nostalgic 1980s AI demo showing a computer being asked to draw a simple yellow circle. The video then cut to modern-day OpenAI offices, where employees interacted with Astra using voice alone, commanding it to transform that same yellow circle into a complex rocket ship, build a full 3D game in minutes, and complete an eBay listing—all without the user ever touching a keyboard.
Astra is scheduled to begin rolling out this Thursday to enterprise customers through OpenAI’s gated access program, known as Daybreak. The company plans to extend access over the coming days to ChatGPT Plus, Pro, Business, and Enterprise users, as well as providing availability via the OpenAI API and major cloud platforms, including AWS Bedrock and Microsoft Azure.
From Chatbots to Digital Operators
The enterprise case for Astra is built entirely on its potential to function as a digital worker. OpenAI notes that the model is capable of executing a wide array of administrative and technical tasks, including filling out online forms, updating CRM records, managing calendars, conducting deep web research, and drafting professional documents. Furthermore, it can manipulate complex spreadsheets, analyze scientific data within Python notebooks, operate within business intelligence platforms like Power BI, and even install or troubleshoot software.
This capability signals a shift in enterprise AI architecture. During the recent generative AI boom, companies have been forced to build intricate "connectors," plugins, and retrieval systems to bridge the gap between their corporate data and AI models. Brockman argued that the arrival of computer-use agents will allow businesses to bypass much of this painstaking integration work. Because modern software is already designed with a human user in mind—complete with visual interfaces and input fields—a sufficiently intelligent agent can interact with these applications directly.
"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. With Astra, an agent can simply "zip through spreadsheets, fill out forms, and navigate across web pages."
Reflecting on the company’s history, Brockman noted that this was the vision from the very beginning. Researchers at OpenAI had long discussed training an agent around the same fundamental inputs and outputs used by humans: pixels, keyboards, and mice. "I feel like we’ve really achieved the first agent that feels like it’s actually able to do that in a way that’s just so extremely useful," he added.
Performance metrics suggest a significant leap in efficiency. On an offline subset of OSWorld 2.0, Astra achieved a score of 72.6% while requiring approximately 40 minutes per task. In comparison, the predecessor model, GPT-5.6 Sol, scored 65.7% but required roughly 75 minutes per task, representing a 47% reduction in time. Beyond mere speed, OpenAI demonstrated the model’s ability to multitask, such as simultaneously creating a 3D game and preparing a legal agreement, suggesting a move away from the "chatterbox" pattern where a human must provide constant, granular instructions.
A Massive Leap in Training Scale
Aidan Clark, a researcher at OpenAI, described the development of Astra as the company’s most significant training endeavor to date. It is the first OpenAI model pretrained using more than 100,000 DBUs at the company’s Stargate infrastructure. Notably, it is also the first model for which previous iterations played a major role in supervising the training process.
"Based on the evaluations we monitor during pre-training, we believe the jump from Sol to Astra represents a larger increase in capabilities than the jump to Sol represented over previous models," Clark stated. The model’s performance is attributed to the combination of large-scale pretraining and reinforcement learning designed to help the system connect information across long, complex task chains.
The reported benchmark results are striking. Astra achieved a 97.6% score on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1, 95.9% on BenchCAD, 96% on GPQA Diamond, and 100% on ExploitBench. Perhaps most notably, it achieved a 98.6% score on ARC-AGI-3. However, this last figure highlights a growing tension in the industry regarding how to measure intelligence.
Defining the AGI Moment
ARC-AGI-3 has emerged as a critical benchmark for testing whether AI can generalize to unfamiliar problems, yet comparing results is complex. OpenAI’s evaluation uses its own "Responses API" harness, which can differ significantly from how other models are configured.
Recent reports from NVIDIA have further complicated the debate. NVIDIA achieved a 100% score on the ARC-AGI-3 public set using an architecture called AVO (Agentic Variation Operators). However, AVO did not rely on a single "smarter" model; it utilized Claude Opus 5, which had a baseline score of roughly 30%. The AVO architecture achieved the perfect score by incorporating persistent memory, tool access, and recovery mechanisms, allowing the agent to maintain progress over long tasks rather than treating every step as an isolated event.
This has sparked intense debate within the AI community regarding whether we are measuring the "intelligence" of the foundation model or the effectiveness of the entire system architecture. Some argue that benchmarks like ARC-AGI-3 are flawed because they restrict context, while others worry that elaborate "harnesses" might simply be masking a model’s lack of true generalization.
For enterprises, these academic nuances may eventually take a backseat to practical results. Companies are buying outcomes, not benchmark rankings. If an agent can reliably investigate a security incident or modify a codebase, the debate over whether that capability stems from neural weights or tool orchestration may matter less than the reliability and auditability of the output.
Brockman acknowledged the ambiguity surrounding the term AGI. "Everyone has a different definition of AGI," he said. "When we started OpenAI, we kind of thought that there was going to be this well-defined moment that everyone would recognize. It hasn’t played out like that. It’s a much more gray, fuzzy thing." Yet, when pressed on whether Astra qualifies, he was firm: "For me personally, I do think we’re there. I think it’s not unreasonable to feel that we are now in the AGI era."
The Economic View of AI Costs
The emergence of autonomous agents is also changing how OpenAI expects customers to view costs. While the company has released standard API pricing for Astra—with standard mode costing $60 per million tokens for output—Brockman suggested that token-based pricing is becoming a poor proxy for value.
"Pricing tokens doesn’t make any sense," Brockman argued. "Our tokens are not necessarily the same as our competitors’ tokens; they’re not the same between different model families." Instead, he advocated for a "price-per-task" model. If a model is more expensive per token but completes a task correctly on the first attempt—avoiding the costs of retries, human oversight, and multiple inference cycles—it is ultimately the more economical choice.
Governance and the Challenge of Autonomy
The very capabilities that make Astra useful to enterprises—its ability to act, change records, and navigate software—also create significant governance challenges. Because an agent can perform actions that carry real-world consequences, OpenAI has emphasized that its understanding of alignment and safety must evolve in tandem with model capabilities.
Following an incident involving Hugging Face, OpenAI paused certain aspects of its frontier training to tighten security, restrict access to research infrastructure, and improve monitoring. This "defense-in-depth" strategy is intended to ensure that as agents become more autonomous, they remain within their authorized boundaries.
For enterprise users, this means that the control surface is no longer just the prompt. CIOs and security leaders will need to implement controls similar to those used for human employees: scoped permissions, audit trails, and real-time monitoring. OpenAI’s Chief Scientist, Jakub Pachocki, stressed that progress in intelligence does not guarantee progress in alignment, and the company remains committed to pausing development if it cannot maintain sufficient confidence in its safety and monitoring systems.
A Critical Cybersecurity Threshold
In the realm of security, Astra has been designated as the first model to reach the "Critical" threshold under OpenAI’s Preparedness Framework. This means the model is capable of identifying previously unknown vulnerabilities and developing exploit chains without continuous human guidance. While this capability is dual-use—potentially aiding both defenders and attackers—OpenAI is initially prioritizing access for trusted defenders responsible for protecting critical digital infrastructure.
Ultimately, whether or not the industry agrees that GPT-6 Astra is the definitive moment for AGI, the impact on the business world will be measured by the scope of work organizations are willing to entrust to it. If Astra succeeds in shifting the professional landscape toward a model of human-supervised autonomy, the transition to the AGI era may be one that is only fully understood in retrospect, as organizations redefine their workflows around the new capabilities of the machine.
