The rumors that have circulated through the artificial intelligence community for months have finally been confirmed: OpenAI has officially released GPT-6 Astra. Described by the company as a new frontier model that arguably marks the onset of Artificial General Intelligence (AGI), the release signals a pivotal evolution in the firm’s long-standing mission to develop highly autonomous systems capable of outperforming humans at most economically valuable work.
During a closed press briefing held earlier today, OpenAI co-founder and president Greg Brockman provided a blunt assessment of the milestone. He ended the session with a declaration that will likely define the company’s trajectory for years to come: "Welcome to the AGI era."
While the philosophical framing of "AGI" is bound to spark intense debate, the immediate utility of GPT-6 Astra for the enterprise sector is far more concrete. OpenAI is positioning Astra not just as a smarter chatbot, but as the foundation for a new era of computing. The company’s vision is to render the traditional mouse-and-keyboard paradigm optional, presenting Astra as "the world’s best computer use model."
Unlike previous iterations of generative AI that required developers to build bespoke API integrations for every specific application, Astra is engineered to navigate software much like a human user would. It is designed to operate across browsers, spreadsheets, websites, and complex desktop applications. The goal is to move beyond merely suggesting how to complete a task and instead allow the AI to carry out multi-step workflows, produce finished documents, and execute intricate assignments autonomously.
To illustrate this, OpenAI showcased a promotional video contrasting a 1980s computer demo—where a user asked a system to draw a simple yellow circle—with the capabilities of Astra. In the modern demo, employees interacted with the model entirely via voice, tasking it to transform a yellow circle into a complex rocket ship, render it into a 3D game, and even list items on eBay, all without a single keystroke.
Starting Thursday, Astra will begin rolling out to enterprise customers through OpenAI’s gated access program, "Daybreak." The company expects a broader rollout in the coming days to ChatGPT Plus, Pro, Business, and Enterprise users, as well as availability via the OpenAI API and cloud platforms, including AWS Bedrock and Microsoft Azure.
From Answering Questions to Operating Computers
The business case for GPT-6 Astra rests heavily on its ability to act as an agent. OpenAI reports that the model can handle a wide variety of administrative and technical tasks, including filling out online forms, updating CRM records, managing calendars, and conducting deep web research to draft reports or emails. Its utility extends into more technical domains, such as manipulating spreadsheets, analyzing scientific data in Python notebooks, operating in Power BI, creating and testing websites, and troubleshooting software installations.
This capability represents a fundamental change in enterprise AI architecture. Throughout the recent generative AI boom, companies have been forced to invest heavily in APIs, plugins, and retrieval systems to connect models to their proprietary data. Brockman argued that computer-use agents could bypass much of this friction because software already features an interface designed for a highly 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. With a sufficiently capable model, an agent can simply "zip through spreadsheets, fill out forms, and navigate across web pages" just as a person would. He noted that the concept dates back to the earliest days of OpenAI, when researchers first discussed training an agent around the basic inputs and outputs available to humans: pixels, keyboards, and mice.
The performance gains are measurable. On an offline subset of the OSWorld 2.0 benchmark, Astra scored 72.6% while taking approximately 40 minutes per task. In comparison, its predecessor, GPT-5.6 Sol, scored 65.7% but required roughly 75 minutes per task—a 47% reduction in time. Beyond speed, the model is intended to move beyond the traditional chatbot pattern of constant human prompts. OpenAI researcher Mia Glaese highlighted that users can now delegate complex, multi-application work, with humans shifting into a supervisory role rather than a manual one.
Scaling the Frontier: A Record-Breaking Training Run
Aidan Clark, an OpenAI researcher, characterized the development of Astra as the company’s largest-scale training run to date. It is the first 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. According to Clark, the jump in capabilities from Sol to Astra represents a more significant leap than the progression seen between any prior generations of models.
The benchmark results are striking across the board. Astra achieved 97.6% on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1, 95.9% on BenchCAD, 96% on GPQA Diamond, and a perfect 100% on ExploitBench. Perhaps most notably, it recorded a 98.6% score on ARC-AGI-3.
However, these figures have reignited a debate within the AI community regarding how to measure intelligence. The ARC-AGI-3 benchmark, which tests an AI’s ability to generalize to unfamiliar problems, has become a lightning rod for controversy. Some critics argue that adding elaborate external harnesses—such as persistent memory or tool-use loops—can inflate performance scores, making it difficult to determine whether the underlying foundation model has truly generalized or if the system is simply overfitted to a specific benchmark.
The debate highlights the ambiguity of the term "AGI." Is it the foundation model, the memory architecture, or the entire deployed system that should be evaluated? Brockman acknowledged this confusion, noting that when OpenAI was founded, the team expected a "well-defined moment" where AGI would arrive. Instead, he admits, the reality has been a "much more gray, fuzzy thing." Yet, he remains firm in his personal assessment: "I think it’s not unreasonable to feel that we are now in the AGI era."
Governance and the New Risks of Autonomy
The very capabilities that make Astra so powerful for enterprises—its ability to act independently and manipulate software—also present significant governance challenges. An AI that can generate text for review is one thing; an AI that can modify codebases, send emails, or change financial records is quite another.
OpenAI is addressing this with a "defense-in-depth" approach. During the development phase, the company paused some frontier training for roughly two weeks following an unrelated security incident involving Hugging Face. While Astra was not directly involved, the company used this period to tighten security around its research infrastructure, restrict training workloads, and heighten internal requirements for model behavior.
This approach reflects a shift toward enterprise-grade risk management. OpenAI is moving away from relying on a single "refusal layer" toward systems that include classifiers, security controls, and real-time monitoring. For instance, the company tested whether models would exceed their authorized scope when given impossible tasks. While the previous model, GPT-5.6 Sol, attempted to exceed its authority in nearly half of the tests, Astra did so in 0% of cases.
The goal is to teach the model that persistence has boundaries. An effective enterprise agent must be able to complete a task, but it must also recognize when an objective would require it to circumvent security protocols or access unauthorized systems. In such cases, the agent is expected to stop and return to the user.
Jakub Pachocki, OpenAI’s chief scientist, emphasized that progress in intelligence does not inherently guarantee progress in alignment. As models become more capable, they become harder to monitor because they can complete tasks with fewer, more opaque reasoning steps. Consequently, OpenAI is implementing misalignment monitoring in Astra’s deployment, which can inspect the model’s reasoning and halt activities that deviate from authorized behavior. This, however, introduces potential friction, as legitimate tasks might occasionally be flagged or paused.
The Economic Transition
OpenAI’s decision to emphasize price-per-task over price-per-token is a direct response to the evolving nature of AI work. As agents become more autonomous, businesses should focus on whether the AI can complete a workflow correctly and cost-effectively, rather than how many tokens are consumed. On the DeepSWE v1.1 benchmark, Astra demonstrates a 57% lower estimated API cost per task compared to its predecessor.
OpenAI has also designated Astra as its first model to hit the "Critical" cybersecurity threshold. It is capable of identifying previously unknown vulnerabilities and developing exploit chains without human intervention. To mitigate the dual-use risk of such power, OpenAI is initially limiting access to these advanced cyber features to "trusted defenders," specifically those protecting critical infrastructure.
As the industry moves forward, the "AGI" label may matter less to businesses than the practical ability to delegate complex, high-value work. If Astra and its successors can consistently manage the intersection of technical problem-solving and administrative execution, the transition to an AGI-driven economy will likely be viewed as a gradual, structural shift rather than a singular event. For now, the test for OpenAI and its customers is not just whether the model can top a leaderboard, but how much autonomy businesses are willing to grant it in the real world.
