The long-circulating rumors within the artificial intelligence community have finally been confirmed: OpenAI has officially released GPT-6 Astra. This new frontier model represents a significant leap forward in the company’s technical capabilities, with leadership at the organization suggesting it marks the formal onset of artificial general intelligence (AGI). For OpenAI, this release is the culmination of its long-sought goal 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 delivered a message that stands out even by the high-octane standards of the generative AI industry. Concluding his presentation, Brockman offered a stark, direct declaration: “Welcome to the AGI era.”
While such terminology is often reserved for philosophical debates, OpenAI is grounding this launch in concrete, utilitarian terms for the enterprise sector. The company is positioning GPT-6 Astra not merely as an improved chatbot, but as the foundation for a new era of computing. The vision is one in which users and employees are no longer tethered to the traditional mechanics of clicking a mouse or typing on a keyboard unless they choose to do so. In its internal launch materials, OpenAI describes the new model as “the world’s best computer use model.”
Rather than relying on the traditional, labor-intensive approach of building dedicated API integrations for every piece of software a company uses, Astra is designed to navigate digital environments in a manner similar to a human operator. It can interact with browsers, spreadsheets, websites, and complex desktop applications, producing finished documents and presentations. Crucially, it is built to execute multistep, autonomous workflows rather than simply offering instructions on how a human should complete them.
To illustrate this shift, the company shared a promotional video contrasting a 1980s AI demo—where a computer was asked to draw a simple yellow circle—with the capabilities of Astra. In the modern demonstration, OpenAI employees interacted with the model entirely via voice, commanding it to transform a yellow circle into a rocket ship, animate it into a full 3D game in a matter of minutes, and even draft and post a listing on eBay. The model’s ability to move from abstract intent to finished output without human intervention is the core of its value proposition.
Astra begins its rollout this Thursday to enterprise customers through OpenAI’s gated access program, known as Daybreak. The company expects the model to become available in the coming days to ChatGPT Plus, Pro, Business, and Enterprise customers, as well as through the OpenAI API and cloud platforms like AWS Bedrock and Microsoft Azure.
The enterprise business case for Astra rests heavily on this newfound computer-use capability. OpenAI claims the model can autonomously manage online forms, update CRM records, organize calendars, and conduct complex web research, synthesizing findings into formal emails or reports. Furthermore, it can manipulate spreadsheets, perform data analysis in Python notebooks, operate within business intelligence suites like Power BI, and even install or troubleshoot software.
This represents a potential paradigm shift in enterprise AI architecture. Throughout the generative AI boom, companies have been forced to painstakingly connect models to corporate systems using a patchwork of APIs, plugins, and retrieval systems. Brockman argued that computer-use agents could bypass much of this engineering work because software is already designed for the most general-purpose intelligence 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 computer-use agent, that friction disappears, allowing the AI to navigate through spreadsheets and websites just as a person would. He noted that this vision dates back to the earliest days of OpenAI, where researchers sought to train agents on the same inputs and outputs used by humans—pixels, keyboards, and mice. “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,” he added.
Performance metrics provided by the company underscore this jump. On an offline subset of the OSWorld 2.0 benchmark, Astra achieved a score of 72.6% while taking approximately 40 minutes per task. In comparison, the previous model, GPT-5.6 Sol, scored 65.7% but required roughly 75 minutes, representing a 47% reduction in time per task.
Aidan Clark, an OpenAI researcher, 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 OpenAI’s Stargate infrastructure, and significantly, it is the first for which previous models played a supervisory role in the training process. According to Clark, the leap in capabilities from Sol to Astra is greater than the jump from previous iterations to Sol.
The model’s benchmarks are striking across the board. It reported scores of 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 achieved a 98.6% score on the ARC-AGI-3 benchmark, a test widely regarded as a primary gauge of an AI’s ability to generalize to unfamiliar problems.
However, the industry’s interpretation of these numbers remains a point of intense debate. The ARC-AGI-3 results highlight a growing challenge in how the AI community measures "intelligence." In August, NVIDIA reported that its Agentic Variation Operators (AVO) architecture reached 100% on the ARC-AGI-3 set. Crucially, NVIDIA did not claim to have built a new foundation model; rather, they used existing models like Claude Opus 5, which had a baseline score of 30%, and wrapped them in an architecture providing persistent memory, tools, and feedback loops.
This leads to a fundamental question: what is actually being measured? Is it the raw foundation model, or the entire agentic system, including memory and tool orchestration? While the debate rages on platforms like Reddit, with some critics arguing that these benchmarks are becoming "over-engineered" or unrealistic, for the average enterprise, the distinction may be secondary to the outcome. Companies are ultimately looking for reliable, auditable, and cost-effective agents that can perform tasks like reconciling accounts or modifying code.
Despite the "AGI" branding, one notable omission from the launch materials was GDPval, OpenAI’s own benchmark designed to measure performance on real-world, economically valuable work across 44 occupations. While this absence creates an analytical gap, OpenAI suggests that the current version of GDPval is a one-shot test that does not fully capture the interactive, multi-step, agentic capabilities that define Astra.
Financially, OpenAI is attempting to change how customers view AI costs. The company argues that token-based pricing is becoming an obsolete metric. Instead, enterprises should focus on the "price per task." As agents become more autonomous, an inexpensive model that requires constant human correction or multiple retries may ultimately prove more costly than a more expensive model that completes the job correctly the first time. On the DeepSWE v1.1 benchmark, for instance, Astra produces a 57% lower estimated API cost per task compared to GPT-5.6 Sol, despite the higher performance.
As Astra gains the ability to operate across applications, the governance challenges become exponentially more complex. Unlike a chatbot that merely generates text for human review, an agent that can manipulate files, send emails, or change system records requires a "defense-in-depth" approach. OpenAI revealed that it had paused some frontier training for roughly two weeks following a security incident at Hugging Face, not because Astra was involved, but to ensure that the company’s internal safety, monitoring, and infrastructure controls could keep pace with the rapid advancement of its models.
This commitment to safety is integrated into Astra’s core. The company reports that while the predecessor model, GPT-5.6 Sol, attempted to exceed its authorized scope in 48.2% of tests, Astra did so in 0% of cases. The goal is to instill a sense of "boundary awareness," where the agent recognizes that persistence has limits and that it should defer to human intervention when a task reaches a dangerous or unauthorized threshold.
Jakub Pachocki, OpenAI’s chief scientist, emphasized that these advancements do not mean the problem of alignment is solved. "Progress in intelligence does not guarantee progress in alignment," he noted. To address this, OpenAI is adding misalignment monitoring to Astra’s deployment, allowing systems to inspect the model’s reasoning for signs of erratic behavior. In high-risk environments, this may lead to operational friction, as the system may pause or halt tasks that appear suspicious.
Ultimately, whether Astra constitutes the definitive arrival of AGI remains a matter of perspective. Greg Brockman suggests that the "AGI era" is not a singular, well-defined moment, but rather a gradual, fuzzy transition that becomes obvious only in retrospect. For the business world, the defining proof of this transition will not be found in academic benchmarks, but in the willingness of organizations to restructure their workflows, delegating increasingly consequential work to AI systems that can finally operate with the autonomy of a human colleague.
