The rumors that have circulated through the tech industry for months have finally been confirmed: OpenAI is officially releasing GPT-6 Astra. This new frontier model represents more than just an incremental upgrade; the company is explicitly framing its launch as the onset of artificial generalized intelligence (AGI). OpenAI’s long-sought objective—to create highly autonomous systems that can outperform humans at the majority of economically valuable work—appears to have reached a critical inflection point.
During a closed press briefing held earlier today, OpenAI co-founder and president Greg Brockman provided a remarkably direct assessment of the model’s arrival. Ending the session with a declarative statement, Brockman signaled a potential shift in the history of computing: "Welcome to the AGI era."
This framing is unusually bold, even by the standards of a company that has spent years pushing the boundaries of generative AI. However, for the enterprise sector, the implications of GPT-6 Astra are far more concrete than philosophical debates about intelligence. OpenAI is positioning the model as the foundation for a new paradigm of computing, one where the traditional reliance on clicking through browser tabs or typing on a keyboard becomes optional. In its advance materials, OpenAI characterizes Astra as "the world’s best computer use model."
Unlike previous iterations of AI that required developers to build custom API integrations for every specific application, Astra is designed to navigate software much like a human would. It interacts directly with browsers, spreadsheets, websites, and desktop applications. It is capable of producing finished documents and presentations, managing multi-step workflows, and navigating complex interfaces rather than simply providing instructions on how a user might perform those tasks themselves.
To illustrate this shift, OpenAI released a promotional video that bridges the gap between the past and the present. It opens with a 1980s AI demo showing a computer drawing a simple yellow circle. The video then transitions to current OpenAI employees interacting with Astra through voice, tasking it with turning that same circle into a complex rocket ship, building a full 3D game in minutes, and managing an eBay listing—all driven by voice input alone.
Astra is scheduled to begin its rollout this Thursday, initially reaching enterprise customers through OpenAI’s gated access program, Daybreak. The company expects the model to become available in the coming days to users of ChatGPT Plus, Pro, Business, and Enterprise, as well as through the OpenAI API and major cloud platforms, including AWS Bedrock and Microsoft Azure.
From Answering Questions to Operating Computers
The enterprise value proposition of Astra rests squarely on its ability to serve as an agentic operator. According to OpenAI, the model can navigate the mundane yet essential tasks that consume human workday hours: filling out online forms, updating CRM records, organizing calendars, conducting web research, and drafting emails. Its technical capabilities are equally broad, ranging from manipulating spreadsheets and analyzing data in Python notebooks to operating specialized engineering software like KiCad and FreeCAD, and even troubleshooting software installations.
This evolution suggests a significant shift in enterprise AI architecture. During the recent generative AI boom, the standard approach involved connecting models to corporate systems via a complex web of APIs, plugins, and retrieval systems. Brockman argues that computer-use agents like Astra may begin to bypass much of this heavy lifting because software is already designed for the most general-purpose intelligence currently available: 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 agent, he noted, the model can simply navigate the UI, "zip through spreadsheets, fill out forms, and navigate across web pages" just as a person would.
The vision is rooted in OpenAI’s earliest research, where the focus was on training agents to interpret the same basic inputs and outputs that humans use: pixels, keyboard strokes, and mouse movements. "I feel like we’ve really achieved the first agent that feels like it’s actually able to do that in a way that’s just so extremely useful," Brockman said.
The performance gains are measurable. In testing on an offline subset of OSWorld 2.0, Astra achieved a success rate of 72.6% while requiring approximately 40 minutes per task. In comparison, the previous model, GPT-5.6 Sol, scored 65.7% while taking about 75 minutes. This represents an approximate 47% reduction in time per task.
Mia Glaese, an OpenAI researcher, emphasized that the goal is to shift the user’s role from prompting AI to supervising it. "With Astra, users have incredible capabilities at their fingertips and can do things that seemed very far away less than a year ago," 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 Training Leap
Aidan Clark, another lead researcher at OpenAI, described Astra as the company’s largest-scale training run to date. It is the first OpenAI model pretrained using over 100,000 DBUs at the company’s Stargate infrastructure. Furthermore, it marks a milestone in the company’s internal methodology, as it is the first model for which previous AI versions played a significant role in supervising the training of their successor.
"Based on the evaluations 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 noted.
Astra’s capabilities are the result of a deliberate combination of large-scale pretraining and reinforcement learning, specifically designed to teach the model how to connect disparate pieces of information to execute long-horizon tasks. The resulting benchmark scores are highly impressive: 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. Additionally, the model scored 98.6% on ARC-AGI-3.
However, the 98.6% score on ARC-AGI-3 highlights a growing debate in the industry regarding how "intelligence" is measured. ARC-AGI is intended to test whether AI can generalize to unfamiliar problems rather than simply mimicking patterns learned during training. While Astra’s score is high, it is not directly comparable to other models due to variations in testing harnesses.
The issue of "harnesses" is central to current AI discourse. For instance, NVIDIA recently reported that its AVO architecture achieved a 100% score on ARC-AGI-3, but this was accomplished by using a foundation model—Claude Opus 5—which had a baseline score of only 30%. The AVO architecture provided the necessary persistent memory and tool-use mechanisms to bridge the gap. This has led some researchers to argue that long-horizon capabilities emerge from the complete agent system rather than the foundation model in isolation.
The debate has even spilled over into public forums like Reddit, where users debate whether benchmarks like ARC-AGI-3—which restrict context retention—actually measure "generalization" or if they are simply tests of an AI’s ability to function within an artificial, constrained environment.
For enterprises, however, these academic distinctions may matter less than the final outcome. "Everyone has a different definition of AGI," Brockman admitted. "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 the Economics of AI
A significant change accompanies the launch of Astra: OpenAI is pushing a shift from "price-per-token" to "price-per-task." While the company has provided a detailed pricing structure for Astra, Brockman argues that token-based pricing is becoming an increasingly poor proxy for the true cost of enterprise AI.
"Pricing tokens doesn’t make any sense," Brockman said. "Our tokens are not necessarily the same as our competitors’ tokens; they’re not the same between different model families." Instead, he suggests that businesses should evaluate the cost of getting a job done. "It’s just about: can you get the thing done for an appropriate cost at appropriate speed?"
This is a critical distinction for CIOs and IT managers. An inexpensive model that requires constant human correction, multiple retries, and high-latency inference steps may prove significantly more expensive in the long run than a higher-priced model that completes the workflow correctly on the first attempt. On the DeepSWE v1.1 benchmark, for example, OpenAI claims its highest-performing configuration beats the previous model while producing an estimated 57% lower API cost per task.
The Governance Challenge of Autonomous Agents
The very capabilities that make Astra useful to the enterprise also create significant governance hurdles. A chatbot that generates text for human review is one thing; an agent that can actively modify records, send emails, and manipulate files across applications is another.
"Even as models can do more things autonomously, we have to be able to trust them more," said Glaese. "Our understanding of alignment and safety has to advance with model capabilities."
To prepare for this, OpenAI revealed that it had paused some frontier training for roughly two weeks following the recent Hugging Face incident. While Astra itself was not involved, the company used that time to tighten security around its infrastructure, restrict data access, and expand monitoring. This move represents a shift toward enterprise-grade risk management. OpenAI is moving away from simple "refusal" layers toward a defense-in-depth strategy that includes classifiers, security controls, and post-deployment threat response.
For higher-risk enterprise users, OpenAI’s monitoring systems will look for suspicious patterns across multiple prompts, identifying when individually innocuous requests might form part of a larger, malicious attack. This implies that organizations must now think about AI governance not just as prompt moderation, but as the management of agent identity and authorization boundaries.
Ultimately, Astra’s success—and the broader goal of AGI—will not be determined by a single benchmark. Instead, the transition will be measured by how many organizations feel comfortable offloading complex, real-world work to an AI system. As Brockman noted, the arrival of AGI may not be a sudden, explosive event, but rather a gradual economic transition that only becomes fully apparent in retrospect. If the past year is any indication, the pace of that transition is only accelerating.
