The rumors that have circulated through the artificial intelligence community for months have finally been confirmed: OpenAI today announced the release of GPT-6 Astra, a new frontier model that the company believes marks the onset of artificial general intelligence (AGI). This milestone represents the culmination of OpenAI’s long-sought mission to develop "highly autonomous systems that outperform humans at most economically valuable work."
In a closed-door press briefing held earlier today, OpenAI co-founder and president Greg Brockman provided a remarkably direct framing of the announcement. Closing the session, he declared, "Welcome to the AGI era." While such terminology is often reserved for academic debate or speculative forecasting, OpenAI is now treating it as an operational reality. For the enterprise sector, however, the significance of Astra is far more pragmatic. The company is positioning the model as the foundation for a new era of computing, one in which the necessity of clicking through interfaces or manually typing on a keyboard could become a thing of the past.
OpenAI describes GPT-6 Astra as "the world’s best computer use model." Unlike previous iterations of AI that required developers to build bespoke API integrations for every specific software application, Astra is designed to navigate digital environments in much the same way a human would. By operating across web browsers, complex spreadsheets, websites, and desktop applications, the model can execute multistep workflows, generate finished documents, and produce presentations, moving beyond the role of a passive assistant to become an active operator.
To illustrate these capabilities, the company released a promotional video contrasting a 1980s AI demo—which showed a computer simply drawing a yellow circle—with the capabilities of Astra. In the modern demonstration, OpenAI employees interacted with the model entirely through voice commands. They instructed Astra to transform a yellow circle into a rocket ship and then into a fully rendered 3D game in a matter of minutes, followed by the creation of a complete product listing on eBay. The demonstration emphasized that these complex actions were carried out solely through natural language interaction.
GPT-6 Astra is scheduled to begin rolling out this Thursday to enterprise customers through OpenAI’s gated access program, Daybreak. Following this initial release, the model will be made available over the coming days to ChatGPT Plus, Pro, Business, and Enterprise customers, as well as through the OpenAI API and major cloud platforms, including AWS Bedrock and Microsoft Azure.
From Answering Questions to Operating Computers
The core of the enterprise value proposition for Astra lies in its ability to handle "computer use" tasks. OpenAI notes that the model is capable of filling out intricate online forms, updating CRM records, managing calendars, conducting deep web research, and drafting results into professional documents or emails. Its utility extends to technical workflows as well: Astra can manipulate spreadsheets, analyze complex scientific data within Python notebooks, work in Power BI, create and test websites, and operate engineering software such as KiCad and FreeCAD.
This shift points toward a fundamental evolution in enterprise AI architecture. During the recent generative AI boom, organizations have been forced to invest heavily in connecting models to corporate systems through APIs, plugins, and custom-built retrieval systems. Greg Brockman argued that the arrival of computer-use agents will allow companies to bypass much of this arduous integration work. Because modern software is already designed for a general-purpose intelligence—the human user—an agent capable of navigating a graphical interface can theoretically interact with these tools without needing a dedicated connector.
Brockman noted that for years, the industry has been bottlenecked by the need for engineers to painstakingly build connections into every tool. With a sufficiently capable agent, that friction is removed. An AI can simply "zip through spreadsheets, fill out forms, and navigate across web pages," essentially acting as a digital employee. This vision returns to the roots of OpenAI’s early research, where the focus was on training agents to handle the same inputs and outputs as a human: pixels, keyboard strokes, and mouse movements. According to Brockman, the company has finally achieved an agent that performs this in a way that is "extremely useful."
Performance metrics appear to support this claim. On an offline subset of the OSWorld 2.0 benchmark, Astra achieved a score of 72.6% while taking 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 speed, the company demonstrated Astra’s ability to perform multiple, disparate tasks—such as building a 3D game and drafting a legal agreement—simultaneously. For enterprises, this move from simply prompting an AI to supervising an autonomous agent represents a shift that may carry more weight than any academic benchmark.
A Landmark Training Effort
Aidan Clark, a researcher at OpenAI, described the development of Astra as the company’s most ambitious training run to date. It is the first OpenAI model pretrained using more than 100,000 DBUs at the company’s Stargate infrastructure. Furthermore, it is the first model where previous iterations played a significant role in supervising the training of the next, a recursive process that seems to have accelerated its development.
Clark suggested that the jump in capabilities from GPT-5.6 Sol to Astra is greater than the improvement seen in any previous generation of OpenAI’s models. The combination of massive-scale pretraining and reinforcement learning has been specifically tuned to teach the model how to connect information and execute long-horizon tasks, where a single error at the start could derail an entire workflow.
Benchmark results for Astra are notable. It recorded scores of 97.6% 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 controversially, it achieved a 98.6% score on ARC-AGI-3, a benchmark designed to measure whether an AI can generalize to unfamiliar problems.
However, this score has reignited a debate within the industry regarding the measurement of AGI. Because ARC-AGI-3 scores can vary significantly depending on the "harness" or the system architecture surrounding the model, the figure is not a direct, apples-to-apples comparison to previous models. For example, NVIDIA recently reported that its AVO architecture achieved a 100% score on ARC-AGI-3, but this was accomplished by using Claude Opus 5—a model that had a baseline performance of only 30%. The performance was boosted by adding persistent memory, tool-use, and recovery mechanisms.
This leads to a critical question for the industry: what exactly is being measured? Is it the foundation model itself, or the entire agentic system? As enterprises prioritize outcomes over "benchmark purity," the distinction may become secondary. If an agent can reliably perform tasks like reconciling accounts or assembling a financial model, the origin of that capability—whether in the neural weights or the surrounding orchestration—matters less than the reliability and auditability of the result. Brockman acknowledged this ambiguity, noting that while the team originally expected a "well-defined moment" for AGI, it has instead proven to be a "gray, fuzzy thing." Nevertheless, he remains firm in his personal assessment that we have crossed the threshold into the AGI era.
Rethinking Costs and Governance
The introduction of Astra also signals a change in how OpenAI wants enterprises to view costs. The company is steering the conversation away from "price-per-token," which Brockman dismissed as a poor proxy for economic value, toward "price-per-task." Because Astra can complete complex workflows more efficiently than its predecessors, OpenAI argues that businesses should focus on the cost of achieving a finished outcome. On the DeepSWE v1.1 benchmark, for example, Astra’s highest-performing configuration achieved better results than GPT-5.6 Sol at a 57% lower estimated API cost per task.
However, the increased autonomy of Astra introduces a complex governance problem. A chatbot that generates text is easily inspected by a human, but an agent that can manipulate files, send information, and take actions across applications requires a different level of oversight. Mia Glaese, an OpenAI researcher, emphasized that as models become more capable, they must also become more "aligned."
In the weeks leading up to the release, OpenAI took proactive steps to ensure safety. The company reportedly paused some frontier training for roughly two weeks following an incident at Hugging Face, not because Astra itself was deemed dangerous, but to ensure that internal security and monitoring controls were keeping pace with the model’s rapidly advancing capabilities. This "defense-in-depth" approach includes classifiers and offline detection systems that monitor for abuse patterns across multiple prompts, rather than just analyzing individual requests.
For enterprise CIOs and security leaders, this means that the control surface for AI governance has expanded. Organizations can no longer rely on simple content filtering. They must now consider sequences of actions, the model’s understanding of its own authorization boundaries, and how to detect suspicious behavior while it is in progress.
Jakub Pachocki, OpenAI’s chief scientist, warned that progress in intelligence does not automatically guarantee progress in alignment. He stressed that observability is becoming a primary infrastructure bottleneck. As models become more intelligent, they may require fewer "reasoning tokens" to reach a conclusion, making it harder for human supervisors to inspect the model’s internal logic. OpenAI has committed to pausing further scaling if it finds that its ability to monitor and align these models is beginning to degrade.
The stakes are highest in the realm of cybersecurity. OpenAI has designated Astra as its first model to meet the "Critical" cybersecurity threshold under its Preparedness Framework, as it is capable of identifying previously unknown vulnerabilities and developing exploit chains without human guidance. To manage this "dual-use" risk, the company is limiting the most advanced cyber capabilities of the model, prioritizing access for trusted defenders and organizations responsible for protecting critical digital infrastructure.
Ultimately, whether or not Astra is definitively "AGI" may matter less to the business world than the observable results it delivers. If organizations begin to restructure their workflows around these agents—allowing AI to execute the intermediate steps of a project while humans provide high-level direction and judgment—the transition to an AI-driven economy will be well underway. As Brockman suggested, the arrival of the AGI era may not be marked by a single, definitive moment, but rather by a gradual shift that becomes undeniable in hindsight. For now, the test for OpenAI and its customers is simple: how much consequential, real-world work will organizations be willing to entrust to these machines?
