The rumors that have circulated through the tech industry for months have finally been confirmed: OpenAI is officially releasing GPT-6 Astra. The company describes this new frontier model as a pivotal milestone, one that it believes likely signals the arrival of artificial generalized intelligence (AGI)—the long-sought objective of developing highly autonomous systems capable of outperforming humans at the vast majority of economically valuable tasks.
During a closed-door press briefing held earlier today, OpenAI co-founder and president Greg Brockman provided a blunt assessment of the launch. Bringing the session to a close, he offered a declaration that underscored the gravity of the release: "Welcome to the AGI era."
Such a bold framing is rare, even by the high-stakes standards of the frontier AI sector. Yet, for the enterprise market, the significance of Astra is rooted in something far more tangible than philosophical debates about intelligence. OpenAI is positioning the model as the foundation for a new paradigm of computing, one where the traditional interface of a mouse, keyboard, and clicking through menus becomes optional. In its advance materials, the company explicitly labels Astra as "the world’s best computer use model."
The operational philosophy behind Astra represents a departure from current AI integration strategies. Historically, companies have been forced to build dedicated API integrations for every specific application an AI might need to interact with. Astra is designed to bypass this cumbersome bottleneck by navigating software much like a human does. It is capable of working across web browsers, complex spreadsheets, websites, and desktop applications. Rather than simply providing instructions to a user on how to complete a task, Astra can produce finished documents, build presentations, and execute multi-step workflows independently.
To demonstrate this, OpenAI shared a promotional video that juxtaposed a nostalgic 1980s-era AI demo—where a computer was asked to draw a simple yellow circle—with modern footage of OpenAI employees interacting with Astra via voice. In the demonstration, employees asked the model to transform that same yellow circle into a rocket ship, then a fully functional 3D game, and finally an active eBay listing, all through fluid, natural-language commands.
Starting this Thursday, the model will be rolled out to enterprise customers through OpenAI’s gated access program, Daybreak. The company expects 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
For the enterprise, the value proposition of Astra rests almost entirely on its capacity for "computer use." The model is engineered to handle a broad spectrum of administrative and technical chores, including filling out online forms, updating CRM records, managing calendars, and conducting deep web research to draft comprehensive reports or emails. Its technical versatility extends to manipulating complex spreadsheets, executing Python code in notebooks, navigating Power BI dashboards, and even troubleshooting or installing software.
This capability represents a fundamental shift in enterprise AI architecture. For the duration of the current generative AI boom, the industry has focused on connecting models to corporate systems through a patchwork of APIs, plugins, and retrieval systems. Greg Brockman argues that these "computer-use agents" will begin to render that painstaking integration work obsolete. Because enterprise software is already designed with the human user in mind, a sufficiently advanced intelligence can interact with those same interfaces directly.
"We’ve been bottlenecked over this entire era by people writing connectors and very painstakingly building these connections into all these tools that people can already use," Brockman noted during the briefing. By allowing an agent to "zip through spreadsheets, fill out forms, and navigate across web pages," the company aims to move beyond the era of simple chat-based assistance. According to OpenAI researchers, Astra is the first agent to successfully replicate that human-like navigation in a way that is profoundly practical for daily professional life.
On internal tests using an offline subset of OSWorld 2.0, Astra achieved a score of 72.6% while taking approximately 40 minutes to complete tasks. This compares favorably to its predecessor, GPT-5.6 Sol, which scored 65.7% while requiring roughly 75 minutes per task—a 47% increase in efficiency. OpenAI suggests that the shift from merely prompting an AI to actively supervising its autonomous work will become the primary focus for businesses looking to gain a competitive edge.
Scaling the Training Frontier
Aidan Clark, a researcher at OpenAI, described the development of Astra as the company’s most ambitious training project to date. It is the first model to be pretrained using more than 100,000 DBUs at the company’s Stargate infrastructure. Furthermore, it marks a milestone in self-improvement for the company, as previous models played a significant role in supervising the training process for Astra.
"Based on the evals 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 said. This leap is attributed to the integration of large-scale pretraining with reinforcement learning techniques designed to help the model maintain context and execute increasingly long, complex workflows.
Astra’s benchmark performance is notable, with reported scores of 97.6% on FrontierMath Tier 4 v2, 95.9% on BenchCAD, 96% on GPQA Diamond, and a perfect 100% on both ExploitBench and ARC-AGI-3.
However, the 98.6% score on ARC-AGI-3 has sparked a debate within the industry regarding how we measure intelligence. ARC-AGI has become a gold standard for testing whether an AI can generalize to unfamiliar problems rather than simply mimicking its training data. While Astra’s score is high, it highlights the ambiguity surrounding "system-level" performance. As evidenced by NVIDIA’s recent work with their AVO architecture—which utilized an underlying model with a 30% baseline to reach a 100% ARC-AGI score through improved memory and feedback loops—long-horizon capabilities often emerge from the entire agentic system rather than the foundation model in isolation.
This ambiguity leads to the central question: what exactly are we measuring? Is it the raw model, the model plus memory, or the entire deployed system? For enterprises, the distinction may eventually fade in importance. Companies are looking for reliable outcomes—whether an agent can reconcile accounts, investigate a security incident, or modify a codebase—and the specific origin of that ability may matter less than its cost, reliability, and auditability.
The Governance Challenge of Autonomy
As Astra moves from generating text to taking action, the governance challenges for organizations increase proportionally. A chatbot that produces a summary for a human to review is a fundamentally different risk profile than an agent capable of modifying records, sending emails, or navigating enterprise software.
"Even as models can do more things autonomously, we have to be able to trust them more," said OpenAI researcher Mia Glaese. "Our understanding of alignment and safety has to advance with model capabilities."
To address these concerns, OpenAI has adopted a defense-in-depth strategy. This includes not just model behavior, but system-level classifiers and offline detection tools that can identify patterns of abuse across multiple interactions. During the development of Astra, OpenAI actually paused certain frontier training for two weeks following the well-publicized Hugging Face incident. While Astra itself was not involved, the company used the pause to tighten security around its research infrastructure and ensure its safety protocols were keeping pace with the model’s capabilities.
This approach reflects a shift toward enterprise-grade risk management. For companies deploying these agents, the focus must move beyond the prompt to a broader understanding of the agent’s authorization boundaries, audit trails, and real-time monitoring. OpenAI’s own testing showed that while earlier models might attempt to exceed their authorized scope when given difficult objectives, Astra’s alignment training ensures it recognizes these boundaries and returns to the user when a task cannot be completed within the permitted constraints.
Price, Performance, and the Future of Work
The pricing model for Astra is also designed to reflect this shift. While OpenAI has provided clear costs per million tokens for standard and fast modes, Greg Brockman emphasized that token-based pricing is becoming a poor proxy for value. Instead, he argued, businesses should evaluate the "price per task." A model that is technically cheaper per token but requires multiple human interventions or retries to complete a workflow is ultimately more expensive than a more capable model that achieves the correct result on the first attempt.
OpenAI’s data suggests that Astra outperforms GPT-5.6 Sol on tasks like software engineering while simultaneously reducing the estimated API cost per task by approximately 57%.
Looking ahead, OpenAI remains committed to the principle that progress in intelligence must be matched by progress in alignment. Chief scientist Jakub Pachocki emphasized that the company is prepared to halt further scaling if it cannot maintain sufficient confidence in its ability to monitor and govern model behavior.
Ultimately, the arrival of GPT-6 Astra suggests that AGI may not be marked by a single, definitive "eureka" moment that everyone recognizes at once. Instead, it may present as a gradual, economic transition that becomes obvious only in hindsight. As organizations begin to restructure their internal workflows—handing off objective-setting to humans and execution to agents—the real test of AGI will not be found on a leaderboard, but in the growing amount of consequential, high-value work that businesses are comfortable delegating to these systems. For now, OpenAI is inviting the world to participate in the first phase of that transition.
