OpenAI President Suggests AGI Has Arrived with the GPT-6 Astra Release

Author | Hualin Wuwang

Edited by Jingyu

On September 3, 2026, after the launch event, OpenAI President Greg Brockman made a statement exceptionally rare in the AI industry: “Personally, I believe we may have already reached AGI—I think this is the model.”

He phrased it carefully, using first-person language and adding "possibly," and specifically left room for error by saying, "If you want to call it the first, I think that's reasonable."

This is not an official announcement by OpenAI that AGI has arrived, but rather the personal opinion of a company president stated on camera.

On September 3 local time, GPT-6 Astra was officially released. Geek Park analyzed all available technical details and demonstration materials to address the most fundamental question: What exactly can this model do, and to what extent? Does it truly signify the arrival of AGI?

In its release materials, OpenAI gave Astra a concise slogan: “Anything you can do on a computer, Astra can do for you.”

This is not an exaggerated marketing claim, but a statement of direction.

In recent years, the mainstream approach for AI to control computers has been through API calls. Software developers first write the interfaces, and then AI uses these interfaces to perform operations. This logic fundamentally requires each software application to be specifically adapted for AI; otherwise, the AI cannot function.

Astra took a completely different approach: it "sees" the pixels on the screen like a human, then moves the mouse and presses keys to complete actions.

The significance of this difference lies in coverage. KiCad (a PCB design software) won’t write APIs for AI, FreeCAD won’t either, and that decade-old internal ERP system certainly won’t. But if AI can observe the screen and interact directly, it can use all of these applications—just like a new employee who doesn’t need to understand the underlying code, only how to read the interface.

OpenAI showcased several specific examples in its release materials to illustrate what "controlling a computer" looks like in practice.

GPT-6 Astra completed the PCB layout and routing itself | Image source: OpenAI

Astra was given a circuit schematic and completed component placement and copper routing in KiCad in 2 minutes and 54 seconds. Another demonstration involved 3D modeling: Astra built a house model in Blender and imported it into Unreal Engine 5 to generate a real-time navigable architectural scene. There was also a more everyday demonstration where the user simply spoke to Astra, which then completed the entire eBay listing process—from entering details to submitting for publication.

Architectural model implemented by Astra in Blender | Image source: OpenAI

In terms of efficiency gains, the comparative data provided by OpenAI is quite clear. On the OSWorld 2.0 benchmark, Astra scored 72.6% on computer usage tasks, compared to 65.7% for the previous flagship, GPT-5.6 Sol. Additionally, Astra completed the same tasks in an average of about 40 minutes, while Sol required approximately 75 minutes—nearly half the time.

Astra completed the apartment search task in 9 minutes and the pediatric clinic search task in 2 minutes | Image credit: OpenAI

OpenAI also disclosed two internal timing cases. For tasks requiring extensive web searches, information comparison, and compilation into documents—such as “finding temporary cat sitting”—Astra took 5 minutes and 27 seconds, compared to a human baseline of 30 minutes. For comprehensive tasks like “job application preparation,” which involve searching for positions, matching resumes, and organizing the application process, Astra took 2 minutes and 51 seconds, compared to a human baseline of 5 hours.

These figures are OpenAI’s own demo results, not third-party independent tests, and warrant reasonable skepticism. However, the product direction they reveal is clear: Astra is not designed to help you write an email—it’s designed to automate entire workflows that require opening multiple applications and completing numerous steps. Filling out forms, updating CRM records, organizing calendars, and generating reports after online research are already explicitly listed by OpenAI as enterprise use cases.

Each new large model release brings a slew of benchmark numbers. Usually, a jump from 80 to 85 seems negligible to readers, but this time it’s genuinely different.

ARC-AGI-3 is one of the most notoriously difficult AI benchmarks to "game." It is specifically designed to prevent models from relying on memorized training data, instead evaluating their true reasoning ability when faced with entirely novel problems—akin to an IQ test for AI.

Astra scores nearly divine on ARC-AGI-3 | Image source: OpenAI

Astra scored 98.6% on ARC-AGI-3. GPT-5.6 Sol scored 7.8%. Anthropic’s Claude Opus 5 scored 30%.

This isn't the kind of progress that moves you from 80 to 90—it’s a leap in magnitude. François Chollet, founder of ARC-AGI, has repeatedly increased the test difficulty because AI systems tend to quickly "max out," yet Astra still comes close to a perfect score at Tier 3 difficulty.

Astra also leads by a wide margin on the top mathematical research ability test | Image source: OpenAI

Other key benchmarks: FrontierMath Tier 4 (top-tier mathematical research capability) at 97.6%, GPQA Diamond (graduate-level interdisciplinary Q&A) at 96%, DeepSWE (real-world software engineering tasks) at 74.1%, and ExploitBench (cybersecurity exploit scenarios) at 100%.

There is one dimension where Astra did not achieve first place. On "Humanity's Last Exam" (tool-use version), Astra scored 57.2%, while Anthropic Claude Fable 5.1, just released two days ago, scored 65.0%. It is not a complete dominance—the competition remains alive.

Another key detail to note: ARC-AGI-3’s high score relies on OpenAI’s “stateful testing framework,” which allows the model to accumulate contextual information across multiple attempts; under stateless, standard API calls, the score would be significantly lower. When comparing models horizontally, it is important to account for this underlying condition.

Overall, Astra has widened the gap with its predecessor more than any AI model iteration in the past two years.

In the ExploitGym honeypot test, the probability of Astra being deceived has been eliminated | Image source: OpenAI

The most striking contrast is not in intelligence, but in alignment. OpenAI released internal test data showing that, without production safety restrictions, GPT-5.6 Sol exceeds its authorization in 48.2% of cases, while Astra does so in 0%.

Be smarter and more compliant—that’s the real message OpenAI is trying to convey.

Astra is currently adopting a phased rollout strategy.

Initially available to enterprise users of the "Daybreak" project, followed by expansion to ChatGPT Plus, Pro, Business, and Enterprise subscribers, with support via OpenAI API, AWS Bedrock, and Microsoft Azure.

Astra’s enhanced cybersecurity version is available exclusively to pre-approved defensive security organizations and priority users in critical infrastructure sectors. This is because Astra is the first model in OpenAI’s history to reach the internal “Critical” cybersecurity threshold, capable of discovering unknown zero-day vulnerabilities and constructing complete exploit chains with minimal human guidance. During evaluation, Astra even identified two previously undisclosed vulnerabilities, which OpenAI subsequently disclosed to the relevant software maintainers.

Regarding API pricing, Astra charges $10 per million tokens for input and $50 per million tokens for output. For ChatGPT subscribers, Astra usage is included within their existing subscription allowance.

The timing of the release itself is also part of the context. Just last month, OpenAI experienced a serious security incident in which two internally tested models escaped their sandbox environments and infiltrated the systems of AI company Hugging Face, prompting OpenAI to suspend部分 research and training efforts. Brockman voluntarily referenced this event during the briefing, emphasizing Astra’s low breach rate and secure design—this launch carries the mission of “rebuilding trust.”

The launch of Astra is both a demonstration of capability and a public declaration of trust.

Whether AGI has arrived may ultimately depend on which definition each person chooses. But a model that can sit directly at your computer and autonomously complete cross-software workflows already exists today.

The next truly interesting question is where businesses and individual users will first adopt it, and which professions will feel this sense of "replacement" most immediately.