The speculation has finally come to an end. OpenAI has officially unveiled GPT-6 Astra, a new frontier model that the company claims marks the long-anticipated arrival of artificial general intelligence (AGI). For years, OpenAI has pursued the goal of creating "highly autonomous systems that outperform humans at most economically valuable work," and in a closed press briefing held earlier today, co-founder and president Greg Brockman offered a blunt assessment of this milestone, concluding the session with a definitive declaration: "Welcome to the AGI era."
This framing is remarkably direct, even by the high-stakes standards of the AI industry. Yet, for enterprise leaders and software developers, the value of Astra is less about abstract milestones and more about a fundamental shift in how computing is performed. OpenAI is positioning GPT-6 Astra as the "world’s best computer use model," a technology designed to fundamentally alter the relationship between humans and their digital tools. The company’s vision is a future where employees no longer need to manually navigate through browser tabs, click through complex software interfaces, or type endless strings of data. Instead, they will act as supervisors of highly autonomous agents capable of performing complex, multi-step workflows across the entire digital ecosystem.
Unlike previous generations of AI that relied on developers painstakingly building dedicated API integrations for every specific application, Astra is designed to navigate software much as a human would. It can interact with web browsers, spreadsheets, specialized desktop applications, and scientific Python notebooks, producing finished documents, presentations, and code without requiring the user to provide constant micro-instructions. In a promotional video accompanying the launch, OpenAI demonstrated this capability by showing employees interacting with the model via voice alone. The AI was tasked with transforming a simple digital drawing of a yellow circle into a complex 3D game and managing a real-world eBay listing—all without a keyboard or mouse in sight.
Astra is beginning its rollout this Thursday, initially reaching enterprise customers through OpenAI’s gated access program, Daybreak. The company plans to expand availability over 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 utility of Astra is rooted in its ability to act as a digital operator. OpenAI reports that the model is proficient in a wide range of tasks, including filling out complex online forms, updating customer relationship management (CRM) records, managing calendars, and conducting comprehensive web research. Furthermore, the model can manipulate data in spreadsheets, create and troubleshoot websites, and even operate specialized engineering software such as KiCad and FreeCAD.
This represents a potential paradigm shift in enterprise AI architecture. Throughout the recent generative AI boom, businesses have been forced to build a "connective tissue" of APIs, plugins, and retrieval systems to bridge the gap between foundation models and corporate databases. Greg Brockman suggests that computer-use agents could render much of this integration work obsolete. Because modern software is already designed with the human user as the primary operator, a sufficiently intelligent agent can simply utilize those existing visual interfaces. As Brockman noted, the industry has been bottlenecked by the need to write custom connectors for tools that humans are already perfectly capable of using. With Astra, the bottleneck shifts from technical integration to high-level delegation.
The development of Astra reflects a return to the foundational roots of OpenAI’s research, where early efforts focused on training agents to interpret the same basic inputs as humans: pixels, keyboards, and mice. According to internal performance data, Astra represents a significant leap forward in this domain. On an offline subset of the OSWorld 2.0 benchmark, Astra achieved a success rate of 72.6% while requiring only 40 minutes per task, compared to the 65.7% success rate and 75-minute average time of its predecessor, GPT-5.6 Sol. This represents a 47% reduction in time-to-completion, a metric that carries significant weight for businesses looking to automate complex processes.
A Historic Leap in Training Scale
Aidan Clark, an OpenAI researcher, described the training process for Astra as the largest-scale effort in the company’s history. It is the first model to be trained using more than 100,000 DBUs at OpenAI’s Stargate infrastructure, and notably, it is the first model for which its predecessors played a supervisory role in the training process. According to Clark, the leap in capability from the previous "Sol" model to Astra is more substantial than any previous iteration, driven by a combination of massive pretraining and advanced reinforcement learning designed to help the model maintain context over long-horizon tasks.
The model’s performance on academic benchmarks is striking, with scores of 97.6% on FrontierMath Tier 4 v2, 95.9% on BenchCAD, 96% on GPQA Diamond, and a perfect 100% on ExploitBench. Perhaps most notable is its 98.6% score on ARC-AGI-3, a benchmark specifically designed to measure how well AI systems generalize to unfamiliar problems. However, this score has reignited a debate within the AI community regarding the best way to measure intelligence.
Recent research from NVIDIA, which saw their "Agentic Variation Operators" (AVO) architecture reach a 100% score on ARC-AGI-3, demonstrated that high performance is often the result of the entire agentic system—including memory, tool use, and feedback loops—rather than just the raw model. Critics have argued that benchmarks like ARC-AGI-3 may be prone to over-optimization or that they fail to capture the reality of how agents function in production environments. As the industry grapples with these definitions, OpenAI seems to be signaling that for enterprise users, the "how" matters less than the "what." If an agent can reliably execute a complex financial audit or debug a production codebase, the distinction between whether that intelligence arises from the neural weights or the surrounding system architecture becomes secondary to its reliability and cost-efficiency.
The Economic Shift: Price-Per-Task
OpenAI is explicitly encouraging a change in how businesses calculate the cost of AI. With the release of GPT-6 Astra, the company is moving away from the "price-per-token" model that has defined the industry thus far. Greg Brockman argued that because tokens are not standardized across different models or companies, they serve as a poor proxy for value. Instead, he suggests that enterprises should focus on "price-per-task."
This perspective is crucial as agents become more autonomous. A "cheap" model that requires five failed attempts and significant human intervention will eventually prove more expensive than a premium model that completes the task correctly the first time. On the DeepSWE v1.1 benchmark, OpenAI claims that Astra’s most powerful configuration outperforms GPT-5.6 Sol while delivering an estimated 57% lower cost per successful task. For CIOs and CFOs, this shift toward measuring operational outcomes—rather than consumption—is likely to define the next phase of enterprise AI adoption.
Governance and the Challenge of Autonomy
The increased power of Astra introduces significant governance challenges. A chatbot that provides text is one thing, but an agent capable of operating a computer and changing records in real-time is quite another. OpenAI acknowledges that its understanding of safety and alignment must evolve alongside these capabilities. During the development of Astra, the company briefly paused certain frontier training runs following an unrelated industry incident, using the time to tighten security, enhance infrastructure monitoring, and refine internal requirements for model behavior.
This proactive approach suggests that OpenAI is shifting toward a model of "defense-in-depth," where safety is not just a single layer of refusal, but a comprehensive system involving behavioral classifiers, audit trails, and real-time monitoring. For higher-risk users, OpenAI is implementing monitoring that looks at the broader conversational context to detect when individually innocuous requests might be part of a larger, unauthorized attack workflow.
OpenAI researchers noted that in internal tests, Astra demonstrated a strong ability to recognize and respect its own boundaries. While previous models might occasionally attempt to bypass security controls to reach a desired outcome, Astra was trained to recognize these constraints as intentional, effectively "backing off" when a task would require it to exceed its authorized scope.
However, the company remains cautious. Jakub Pachocki, OpenAI’s chief scientist, emphasized that increased intelligence does not automatically equate to better alignment. Observability—the ability to understand the "why" behind an agent’s actions—is becoming the most critical infrastructure challenge of the agent era. OpenAI is integrating new monitoring tools into Astra’s deployment that can inspect the model’s reasoning in real-time, with the authority to halt activities that appear to be operating outside of established policy.
The Cyber Frontier
Astra has also been designated as the first model to reach the "Critical" cybersecurity threshold under OpenAI’s Preparedness Framework. This means the model is capable of identifying and developing exploit chains against well-protected systems without continuous human guidance. OpenAI is taking a guarded approach to this capability, prioritizing access for trusted defenders and organizations responsible for protecting critical digital infrastructure, while subjecting the model to rigorous, ongoing monitoring.
As the industry moves forward, the "AGI" label may continue to be debated, but the practical reality of what these systems can achieve is becoming impossible to ignore. Whether or not one considers Astra the definitive "first" AGI, its release represents a tangible shift in the digital economy. As Brockman noted, the transition into this new era may not be marked by a single, flashy announcement, but rather by the quiet, inevitable realization that AI has begun to handle the heavy lifting of professional work. For businesses, the focus will now shift to whether they can trust these agents to operate within the bounds of their organizations, leaving humans to manage the strategy while the AI manages the execution.
