The long-standing whispers from the corridors of San Francisco have finally materialized into a concrete, and potentially industry-altering, reality. OpenAI today officially released GPT-6 Astra, a new frontier model that the company asserts marks the onset of artificial general intelligence (AGI). For years, OpenAI has maintained a mission to build "highly autonomous systems that outperform humans at most economically valuable work," and in a closed press briefing held earlier today, the company’s leadership signaled that this threshold has finally been crossed.
Closing the session with a definitive statement, OpenAI co-founder and president Greg Brockman left little room for ambiguity: "Welcome to the AGI era."
This framing is, by any measure, a consequential declaration in an industry defined by hyperbole. However, for the enterprise sector, the immediate significance of Astra is more tangible than a philosophical debate over the definition of intelligence. OpenAI is positioning GPT-6 Astra as a fundamental shift in the nature of computing itself—a world where employees may no longer need to rely on the traditional mouse-and-keyboard paradigm to complete complex digital tasks.
In its promotional materials, OpenAI characterizes the model as the world’s most capable "computer use" system. Unlike previous generations of AI that required developers to painstakingly build dedicated API integrations for every piece of software an agent might need to touch, Astra is designed to navigate software interfaces much like a human does. It can work across browsers, spreadsheets, websites, and desktop applications, producing finished documents, creating presentations, and executing multi-step workflows rather than simply offering instructions on how to perform them.
To illustrate this shift, OpenAI showcased a promotional video contrasting a 1980s demo—where a user asked a computer to draw a simple yellow circle—with a modern demonstration of Astra. In the new clip, OpenAI employees interact with the model entirely through voice, requesting that it transform a simple yellow circle into a rocket ship, then into a fully rendered 3D game, and finally into a functional eBay listing, all in a matter of minutes.
Astra is scheduled to begin its rollout this Thursday to enterprise customers through OpenAI’s gated "Daybreak" program. In the coming days, the model will see broader availability for ChatGPT Plus, Pro, Business, and Enterprise customers, as well as through the OpenAI API and cloud platforms, including AWS Bedrock and Microsoft Azure.
From Answering Questions to Operating Computers
The core of the enterprise value proposition for Astra rests on its ability to act as an agent. OpenAI notes that the model is capable of navigating the full spectrum of modern digital work: filling out online forms, updating CRM records, managing calendars, conducting web research, and drafting results into documents or emails. Beyond basic administrative tasks, it can manipulate complex spreadsheets, analyze scientific data in Python notebooks, operate business intelligence software like Power BI, and even troubleshoot or install software.
This capability points toward a tectonic shift in enterprise AI architecture. Throughout the recent generative AI boom, the primary bottleneck for businesses has been the need to connect models to corporate systems via custom APIs, plugins, and retrieval systems. Greg Brockman argued that the arrival of computer-use agents effectively bypasses much of this integration drudgery, as software already exposes an interface designed for a human user—an interface that a sufficiently intelligent model can now read and operate directly.
"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 the advent of Astra, he added, an agent can simply "zip through spreadsheets, fill out forms, and navigate across web pages."
This goal has been part of OpenAI’s vision since its inception. Researchers long debated the possibility of training an agent using the same basic sensory inputs as a human: pixels, keyboard strokes, and mouse movements. According to Brockman, the company has finally achieved the first iteration of an agent that can perform these tasks with a level of utility that feels genuinely transformative.
The performance gains are reflected in the company’s internal testing. On an offline subset of OSWorld 2.0, Astra achieved a score of 72.6% while requiring roughly 40 minutes per task. In comparison, the previous model, GPT-5.6 Sol, scored 65.7% but required approximately 75 minutes, representing a 47% reduction in time per task. During demonstrations, Astra showed it could manage diverse, complex tasks—such as building a 3D game while simultaneously preparing a legal agreement—without breaking its stride.
OpenAI researcher Mia Glaese noted that the model is designed to move beyond the "chatbot" pattern, where a human is constantly required to provide the next instruction. "With Astra, users have incredible capabilities at their fingertips," Glaese said. "We expect people to delegate much more complex work across applications, with humans directing the work at a much higher level."
A Massive Leap in Training Scale
Aidan Clark, a researcher at OpenAI, described 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. Furthermore, it is the first instance where previous, smaller models played a significant role in supervising the training of their successor.
"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 stated. The model’s performance is attributed to a combination of massive-scale pretraining and reinforcement learning, which specifically teaches the system to connect disparate pieces of information and execute long-horizon tasks.
The benchmark results are impressive, with Astra scoring 97.6% on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1, and 95.9% on BenchCAD. It also reported a 96% score on GPQA Diamond, 100% on ExploitBench, and 98.6% on the highly scrutinized ARC-AGI-3 benchmark.
The Debate Over AGI Benchmarks
The 98.6% score on ARC-AGI-3 has ignited a broader industry conversation about how to measure intelligence. ARC-AGI is a benchmark designed to test an AI’s ability to generalize to unfamiliar problems rather than simply reproducing memorized training data. While Astra’s score is high, comparisons are difficult because the surrounding architecture of an AI system—not just the model itself—can heavily influence the outcome.
Earlier this year, NVIDIA reported that its Agentic Variation Operators (AVO) architecture reached a 100% score on the same benchmark, but it achieved this using Claude Opus 5 as the underlying engine. The improvement was largely driven by mechanisms like persistent memory and feedback loops. This has led some to argue that long-horizon capabilities are an emergent property of the entire system rather than the foundation model alone.
As the industry grapples with these nuances, the distinction between a foundation model, a model with memory, and a fully deployed agent system becomes increasingly blurred. For enterprises, however, the practical reality is that they are purchasing outcomes, not benchmark scores. Whether an agent’s success in reconciling accounts or drafting code stems from neural weights or tool orchestration is less important than its cost, reliability, and auditability.
When asked if Astra qualifies as AGI, Brockman acknowledged the "gray, fuzzy" nature of the term. "When we started OpenAI, we kind of thought that there was going to be this well-defined moment that everyone would recognize," he noted. "It hasn’t played out like that." However, regarding Astra, he was more definitive: "For me personally, I do think we’re there. I think it’s not unreasonable to feel that we are now in the AGI era."
Cost, Governance, and the Future of Work
A notable omission from the launch materials was the company’s internal benchmark, GDPval, which is designed to measure performance on real-world economic tasks across various industries. While Astra’s omission of this metric leaves a gap in the analysis, the model’s performance on other professional benchmarks suggests a significant leap in capability.
OpenAI is also shifting how customers should think about costs. Brockman argued that "price per token" is an outdated metric for enterprise users. Instead, businesses should focus on "price per task." By this metric, OpenAI suggests that Astra is significantly more efficient than its predecessors, as it can complete complex tasks in fewer steps, thereby reducing the total cost of inference.
Yet, this autonomy presents a substantial governance problem. A chatbot is generally passive, but an agent that can operate a computer and interact with production systems is inherently more risky. To address this, OpenAI has implemented a defense-in-depth security strategy. Following a period of heightened internal monitoring, the company has integrated robust cybersecurity safeguards, including system-level classifiers and offline detection systems, to prevent agents from exceeding their authorized scope.
OpenAI chief scientist Jakub Pachocki emphasized that intelligence and alignment are not the same thing. "Progress in intelligence does not guarantee progress in alignment," he warned. Consequently, the company is implementing "misalignment monitoring," which can halt an activity if the model appears to be acting outside of its granted authority.
As Astra is the first model to reach OpenAI’s "Critical" cybersecurity threshold, the company is limiting its most advanced features, particularly for those involved in protecting critical infrastructure. The goal is to ensure that while the model is capable of identifying zero-day vulnerabilities, it remains a tool for defense rather than a weapon for exploitation.
Ultimately, the arrival of GPT-6 Astra suggests that AGI may not be marked by a single, dramatic event, but by a gradual, profound economic transition. If organizations begin to restructure their workflows around these autonomous systems—with humans acting as supervisors rather than operators—the era of AGI will have arrived in practice, even if the debate over the label continues. For now, the focus for CIOs and business leaders will likely shift to a new, practical question: how much responsibility can they safely delegate to an agent that never sleeps, never forgets, and is increasingly capable of doing the work itself?
