The rumors that have circulated through the artificial intelligence community for months have finally been confirmed: OpenAI today officially announced the release of GPT-6 Astra. The company characterizes this new frontier model as a pivotal milestone, one that it believes likely marks the onset of artificial generalized intelligence (AGI)—the long-stated ambition of creating highly autonomous systems capable of outperforming humans across the vast majority of economically valuable work.
During a closed-door press briefing held earlier today, OpenAI co-founder and president Greg Brockman provided a remarkably direct framing of the release. He concluded the session with a statement that signaled the company’s confidence in the magnitude of this launch: “Welcome to the AGI era.”
While the AGI framing is undoubtedly consequential, the more immediate utility of GPT-6 Astra for the enterprise world is considerably more grounded and transformative. OpenAI is positioning Astra not merely as an advanced chatbot, but as the engine for a new era of computing—one where the traditional reliance on clicking a mouse or typing on a keyboard could become entirely optional. The company’s internal documentation, shared in advance with industry analysts, explicitly labels Astra as “the world’s best computer use model.”
Rather than forcing developers to build fragmented API integrations for every individual application within a corporate stack, Astra is architected to navigate software environments in the same way a human would. It is designed to move fluidly across web browsers, spreadsheets, specialized desktop applications, and collaborative suites. The model can produce finished documents, craft complex presentations, and execute multi-step workflows end-to-end, rather than simply providing instructions to a human on how to complete those tasks.
To illustrate these capabilities, OpenAI shared a promotional video contrasting the past and present. The clip begins with a vintage 1980s-era AI demonstration, showing a computer drawing a simple yellow circle based on a prompt. It then cuts to the present day, where OpenAI employees interact with Astra through voice commands alone. In the demonstration, the model is asked to transform that same yellow circle into a detailed rocket ship, then into a fully functional 3D game, and finally to create an active listing on eBay—all through spoken input, without a single manual keystroke.
The rollout begins this Thursday for enterprise customers via OpenAI’s gated "Daybreak" program. In the coming days, the company expects to expand access to ChatGPT Plus, Pro, Business, and Enterprise users, as well as providing availability through the OpenAI API and major cloud infrastructure platforms, including AWS Bedrock and Microsoft Azure.
From Answering Questions to Operating Computers
The primary value proposition for enterprises rests on this shift toward agentic computer use. OpenAI asserts that Astra is capable of filling out complex online forms, updating customer relationship management (CRM) systems, managing intricate calendars, conducting comprehensive web research, and drafting the findings into professional documents or emails. Beyond general administrative tasks, the model demonstrates proficiency in technical environments, including manipulating spreadsheets, analyzing scientific data in Python notebooks, working within Power BI, creating and testing websites, and operating specialized engineering software like KiCad and FreeCAD.
This represents a potential paradigm shift in enterprise AI architecture. Throughout the initial generative AI boom, corporations have been forced to build custom bridges between models and their internal systems using APIs, plugins, and complex retrieval systems. Brockman argued that the arrival of computer-use agents allows organizations to bypass much of this "connective tissue" labor. Because existing enterprise software already exposes an interface designed for human intelligence—the screen, the keyboard, and the cursor—a truly general-purpose AI can simply inhabit that space.
Brockman noted that the industry has been bottlenecked by the painstaking process of writing connectors for tools that humans are already perfectly capable of using. With sufficiently capable computer-use agents, that bottleneck effectively vanishes, allowing the system to zip through spreadsheets and navigate web pages as a human operator would. He reflected on the early days of OpenAI, where researchers first discussed training an agent on the fundamental inputs of human-computer interaction: pixels, mouse coordinates, and keystrokes. He believes Astra is the first model to achieve that goal in a way that is genuinely and extremely useful.
Performance data supports this evolution. On an offline subset of OSWorld 2.0, Astra achieved a 72.6% success rate, completing tasks in an average of 40 minutes. This compares favorably to its predecessor, GPT-5.6 Sol, which scored 65.7% while requiring approximately 75 minutes per task, representing a roughly 47% reduction in time-to-completion.
A Significant Leap in Training Scale
Aidan Clark, an OpenAI researcher, described Astra as the company’s largest-scale training run to date. It is the first model to be pretrained using more than 100,000 DBUs at the company’s Stargate infrastructure. Furthermore, it is the first model for which previous iterations played a primary role in supervising the training of the next generation. Clark emphasized that the leap from Sol to Astra represents a larger increase in capability than the transition to Sol from previous models.
The model’s performance is bolstered by a combination of large-scale pretraining and reinforcement learning, specifically tuned to help the model connect information across disparate sources and execute increasingly long, complex tasks. Benchmarking scores are striking, with Astra reaching 97.6% on FrontierMath Tier 4 v2, 95.9% on BenchCAD, 96% on GPQA Diamond, and a perfect 100% on ExploitBench.
Perhaps most discussed is the 98.6% score on ARC-AGI-3. While this is a high-water mark, it has ignited a debate within the AI community regarding how to measure intelligence. The distinction between a foundation model and a complete agentic system is becoming blurred. Critics and researchers note that other configurations—such as NVIDIA’s AVO architecture—have achieved perfect scores on similar benchmarks by utilizing external memory, tool use, and feedback loops rather than relying solely on the underlying model’s weights. This raises a fundamental question: when we speak of AGI, are we measuring the model itself, or the entire system, including the software harnesses that allow it to persist?
For the enterprise, the answer may be secondary to the results. Companies are seeking outcomes, not benchmark purity. If an agent can reliably reconcile accounts, investigate a security incident, or modify a codebase, the debate over whether that capability stems from neural architecture or external orchestration is far less important than the reliability, cost, and auditability of the output.
The Evolution of Enterprise Costs and Governance
Brockman believes the market is beginning to recognize that traditional token-based pricing is an inadequate metric for AI utility. Because tokens vary between models and families, they do not accurately reflect the economic value of a task. Instead, he suggests that businesses should evaluate "price per task." By this metric, Astra demonstrates significant efficiency; its highest-performing configuration outperforms GPT-5.6 Sol while reducing the estimated API cost per task by approximately 57%.
However, as these agents gain autonomy, they present a more complex governance problem. A chatbot that generates text is limited to the output on a screen, but an agent capable of operating a computer can manipulate files, send information, and take actions across a network. OpenAI acknowledges this, noting that as agents become more capable, they must also become more "aligned."
The company’s approach to safety is evolving into something akin to enterprise risk management. During the development of Astra, OpenAI implemented a series of rigorous security protocols, even pausing certain training runs to ensure that safety and monitoring infrastructure kept pace with the model’s rapidly advancing capabilities. This includes "defense-in-depth" strategies, such as system-level classifiers and offline detection that can identify malicious patterns across multiple prompts—a necessity for preventing agents from being subverted by complex, multi-step attacks.
OpenAI researcher Mia Glaese highlighted that the shift from "prompting" to "supervising" is the true hallmark of the new era. Users will no longer simply issue instructions; they will delegate work and monitor the agent’s trajectory. This necessitates a new set of enterprise controls, including scoped permissions, audit trails, and real-time monitoring.
Chief scientist Jakub Pachocki emphasized that while Astra is the company’s most aligned model to date, progress in intelligence does not inherently guarantee progress in safety. Consequently, OpenAI has pledged to treat monitorability as a hard constraint. If the company cannot maintain sufficient visibility into an agent’s reasoning, it will pause development, regardless of the potential for further intelligence gains.
The Path Forward
The designation of Astra as the first model to reach the "Critical" cybersecurity threshold under OpenAI’s Preparedness Framework highlights the dual-use nature of this technology. While the model can autonomously identify zero-day vulnerabilities, this capability is being restricted initially to trusted defenders to ensure it is used to patch, rather than exploit, critical infrastructure.
As OpenAI moves forward, the "AGI" label remains a subject of ongoing debate. However, for organizations, the definition is becoming increasingly practical. If the industry continues to move toward a reality where humans define goals and AI systems execute the intermediate steps, the arrival of AGI may not be marked by a single, dramatic test result. Instead, it will likely be recognized as a gradual, fundamental shift in the global economy—one that becomes obvious only in retrospect. As Brockman suggested, it is becoming increasingly difficult to argue that we have not already crossed the threshold into this new era. The true test will be the extent to which enterprises are willing to entrust these autonomous agents with the consequential, real-world work that keeps the modern economy running.
