The rumors that have swirled through the technology industry 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 tangible onset of artificial generalized intelligence (AGI). For years, OpenAI has pursued the singular goal of creating "highly autonomous systems that outperform humans at most economically valuable work," and in a closed press briefing held earlier today, the company’s co-founder and president, Greg Brockman, offered an unusually direct assessment of the milestone. He concluded the session with a simple, declarative statement: "Welcome to the AGI era."
Such a bold framing is rare, even by the high-stakes standards of the frontier AI sector. However, for enterprise leaders, the significance of Astra is less about philosophical definitions of machine intelligence and more about a fundamental shift in computing. OpenAI is positioning GPT-6 Astra as the world’s most capable "computer use model," a tool designed to liberate employees from the repetitive cycle of clicking, typing, and manual navigation. By functioning as a digital agent that can navigate software—moving across browsers, spreadsheets, specialized desktop applications, and websites—Astra aims to complete multistep workflows rather than merely instructing users on how to finish them.
From Chatbots to Autonomous Operators
The promotional materials released alongside the model illustrate this leap in capability. In a demonstration video, OpenAI juxtaposed a 1980s AI demo—where a user asked a computer to draw a simple yellow circle—with modern footage of OpenAI employees interacting with Astra. In the new clips, users speak naturally to the model, asking it to transform that circle into a rocket ship, generate a full 3D game in minutes, or even manage complex tasks like creating an eBay listing, all through voice commands alone.
This ability to manipulate software interfaces represents a significant change in enterprise AI architecture. For much of the recent generative AI boom, companies have been forced to invest heavily in "glue" work—building APIs, custom plugins, and retrieval systems to connect models to their specific corporate tools. Brockman argued that computer-use agents like Astra could bypass much of this friction because modern software is already built with an interface designed for a 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. By allowing an agent to perceive and act upon the same inputs a human does—pixels, keyboard inputs, and mouse movements—OpenAI believes it has achieved the first agent that is truly and broadly useful. In internal testing on an offline subset of OSWorld 2.0, Astra achieved a score of 72.6% while taking significantly less time per task than its predecessor, GPT-5.6 Sol, demonstrating a 47% increase in efficiency.
Astra is scheduled to begin rolling out this Thursday to enterprise customers through OpenAI’s gated "Daybreak" program. In the coming days, the model will also reach ChatGPT Plus, Pro, Business, and Enterprise customers, alongside integration availability via the OpenAI API and major cloud platforms like AWS Bedrock and Microsoft Azure.
A Historic Scale of Training
According to Aidan Clark, an OpenAI researcher who spoke during the briefing, Astra represents the company’s most ambitious training run to date. It is the first model pretrained using more than 100,000 DBUs at OpenAI’s Stargate infrastructure. Notably, it is also the first model for which previous iterations played a supervisory role in the training of the next, creating a recursive improvement loop. Clark emphasized that the leap from the previous model, Sol, to Astra represents a larger increase in raw capabilities than any prior generational jump in the company’s history.
The benchmark results are impressive: Astra scored 97.6% on FrontierMath Tier 4, 74.1% on DeepSWE, 95.9% on BenchCAD, and 100% on ExploitBench. Perhaps most notably, it achieved a 98.6% score on the ARC-AGI-3 benchmark, a test widely regarded as one of the most rigorous measures of whether an AI can generalize to unfamiliar problems rather than simply mimicking training data.
However, these numbers have reignited a fierce debate within the AI community regarding how to measure intelligence. The distinction between a foundation model and a complete agent system is becoming increasingly blurred. For instance, NVIDIA recently reported that its AVO architecture achieved a 100% score on ARC-AGI-3, but this was accomplished by adding layers of persistent memory and feedback loops to a base model that performed at only 30% without those enhancements.
This leads to a critical question: what exactly is the object being measured? Is it the neural weights of the model itself, or the entire system—including memory, tool orchestration, and browser access—that surrounds it? OpenAI’s leadership seems to suggest that for businesses, this distinction is secondary to results. "Everyone has a different definition of AGI," Brockman noted. "When we started OpenAI, we kind of thought that there was going to be this well-defined moment that everyone would recognize. It hasn’t played out like that. It’s a much more gray, fuzzy thing."
Redefining Enterprise Economics
The shift toward autonomous agents also fundamentally changes how enterprises should view the cost of AI. OpenAI’s pricing structure for Astra is higher than previous models, with Standard mode priced at $60 per million tokens and Fast mode at $120 per million tokens. Yet, Brockman argues that token pricing is becoming an obsolete metric for measuring the value of AI.
"Pricing tokens doesn’t make any sense," he said. "Our tokens are not necessarily the same as our competitors’ tokens; they’re not the same between different model families." Instead, OpenAI is encouraging businesses to evaluate the "price per task." If an agent can reliably reconcile an account or modify a codebase in one go, the cost of the underlying tokens matters far less than the reliability and speed of the outcome. In benchmarks like DeepSWE v1.1, OpenAI claims its highest-performing configuration reduces the estimated API cost per task by approximately 57% compared to the previous model, despite the higher per-token price.
Governance in the Age of Autonomy
The same capabilities that make Astra a powerful business tool also present a significant governance challenge. Unlike a chatbot, which generates text for a human to review, an agent operating a computer can manipulate files, send emails, and alter records. Recognizing this, OpenAI has emphasized that its alignment and safety research must evolve alongside its model capabilities.
Following a brief, precautionary pause in certain training workloads—a decision prompted by the need to ensure that security and monitoring infrastructure kept pace with model development—OpenAI has implemented a defense-in-depth approach for Astra. This includes cybersecurity safeguards that combine model-level refusals with system-level classifiers and offline detection, designed to identify malicious patterns that unfold over multiple steps rather than in a single interaction.
For high-risk environments, OpenAI is deploying monitoring tools that can inspect the model’s reasoning for signs of unauthorized behavior, with the capability to halt activities that exceed defined boundaries. This is a critical development for enterprise security teams, who must now move beyond simple content filtering to managing "AI identities" that possess scoped permissions, audit trails, and real-time monitoring.
The New Frontier of Cybersecurity
OpenAI has designated Astra as the first model to reach its "Critical" cybersecurity threshold under the company’s Preparedness Framework. This means that, when equipped with the right tools, the model is capable of identifying previously unknown vulnerabilities and developing exploit chains across protected systems without continuous human guidance. While Astra has demonstrated the ability to discover zero-day vulnerabilities, OpenAI is taking a cautious approach to distribution, prioritizing access for trusted defenders and organizations responsible for protecting critical infrastructure.
As OpenAI moves forward, the company has made it clear that safety remains a primary constraint on development. Jakub Pachocki, OpenAI’s chief scientist, noted that the company is prepared to slow down or even halt scaling if it cannot guarantee that its ability to monitor and align these increasingly autonomous systems remains sufficient.
Ultimately, the arrival of GPT-6 Astra suggests that the transition to AGI may not be marked by a single, climactic test result. Instead, it appears to be unfolding as a gradual economic transformation. The real test for Astra, and for the organizations that adopt it, will be the degree to which they are willing to shift their workflows: moving from prompting AI to supervising it, and entrusting increasingly consequential, real-world tasks to machines that can act, reason, and navigate the digital world on their own. As Brockman suggested, while it may be difficult to pinpoint the exact moment the era began, it is becoming increasingly difficult to argue that we have not already arrived.
