TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
GPT-6 Leaked, OpenAI CEO Says: AGI Is Here

GPT-6 Leaked, OpenAI CEO Says: AGI Is Here

2026.09.04
Share
Ability ascension.

Author | Hualin Wuwang

Editor | Jing Yu

On September 3, 2026, OpenAI CEO Greg Brockman said something extremely rare in the AI industry after the press conference: "Personally, I believe we may have already reached AGI. I think it is this model."

He phrased his remarks very cautiously, using the first person, adding "possibly," and leaving room for maneuver by saying, "If you want to say this is the first one, I think that's reasonable."

This was not an official announcement from OpenAI declaring the arrival of AGI, but rather a company president voicing his own judgment on camera.

Local time September 3 saw the official release of GPT-6 Astra. Geek Park deconstructed all available technical details and demonstration materials, attempting to answer the most core question: what exactly can this model do, and to what extent? Does it truly represent the arrival of AGI?

“Computer Control” Approaches Mythical Status

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

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

In recent years, the mainstream path for AI to control computers has been calling APIs. Software developers write the interfaces first, and the AI completes operations through those interfaces. This logic essentially requires every piece of software to be specifically adapted for AI; otherwise, the AI is powerless.

Astra takes a completely different path: like a human, it directly “reads” the pixels on the screen, then moves the mouse and types on the keyboard to complete operations.

The significance of this difference lies in its coverage scope. KiCad (a printed circuit board design software) won’t write APIs for AI, FreeCAD won’t, and neither will an internal legacy ERP system used for ten years. But if an AI can look at the screen and operate directly, it can use all these software applications—just like a brand-new employee who doesn’t need to understand the underlying code, only how to navigate the interface.

In the release materials, OpenAI showcased several specific cases to give a feel for what deployed “computer control” looks like.

GPT-6 Astra independently completed PCB layout and routing | Image credit: OpenAI

Given a circuit schematic, Astra completed component placement and copper trace routing in KiCad, taking 2 minutes and 54 seconds for the entire process. Another demonstration involved 3D modeling: Astra built a house model in Blender, imported it into Unreal Engine 5, and generated an architectural scene capable of real-time walkthroughs. In a more everyday demo, a user simply spoke to Astra, which then executed a full eBay listing workflow, from filling out information to submitting the publication.

Architectural model created by Astra in Blender | Image credit: OpenAI

In terms of efficiency improvements, the comparative data provided by OpenAI is quite intuitive. On the OSWorld 2.0 benchmark, Astra scored 72.6% on computer-use tasks, compared to the previous-generation flagship GPT-5.6 Sol’s 65.7%; while completing the same tasks, Astra required an average of only about 40 minutes, whereas Sol took about 75 minutes, making it nearly twice as fast.

https://www.geekpark.net/news/369800https://www.geekpark.net/news/369800

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

OpenAI also released two internally timed case studies. For tasks like “finding temporary cat care,” which require extensive web searches, cross-referencing information, and compiling documents into a report, Astra took 5 minutes and 27 seconds, with OpenAI citing a human baseline of 30 minutes. For comprehensive tasks like “job application preparation,” involving searching for positions, matching resumes, and organizing the application process, Astra took 2 minutes and 51 seconds, with a human baseline of 5 hours.

These two figures are OpenAI’s own demonstration results, not independent third-party tests, so reasonable skepticism is warranted. However, they reveal a clear product direction: Astra was not designed to help you draft an email; it was designed to complete end-to-end workflows for you that typically involve “opening multiple applications and clicking through numerous steps.” Filling out forms, updating CRM records, managing calendars, and conducting online research to generate reports are already explicitly listed by OpenAI as enterprise application scenarios.

Capabilities “Next Level”

Every large model release brings a barrage of benchmark numbers. Most of the time, going from 80 to 85 seems negligible to readers, but this time it is genuinely different.

ARC-AGI-3 is currently recognized as one of the hardest AI benchmarks to “game.” Its design philosophy specifically prevents models from relying on memorized training data, measuring their true reasoning ability when faced with novel problems—somewhat akin to an IQ test for AI.

Astra’s score on ARC-AGI-3 is nearly divine | Image credit: 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” where you go from 80 to 90 points; it’s a magnitude leap. ARC-AGI creator François Chollet has repeatedly increased the difficulty because AI tends to quickly “max out,” yet Astra remains near perfect on Tier 3 difficulty.

Astra also maintains an unbridgeable lead in top-tier mathematical research capability tests | Image credit: OpenAI

Performance on other key benchmarks: FrontierMath Tier 4 (top-tier mathematical research) 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 generation) at 100%.

There is one dimension where Astra did not take first place. On “Humanity’s Last Exam” (with tool invocation), Astra scored 57.2%, while Anthropic’s Claude Fable 5.1, released just two days prior, scored 65.0%. It is not a total domination; competition still exists.

Another crucial detail must be noted. Astra’s high score on ARC-AGI-3 relies on OpenAI’s “stateful testing framework,” which allows the model to accumulate contextual information across multiple attempts; under stateless, standard API call conditions, the score would be significantly lower. Cross-model horizontal comparisons need to account for this precondition.

Overall, the gap Astra has widened with the previous generation of models is rare in AI iterations over the past two years.

In ExploitGym honeypot tests, Astra’s chances of being deceived are effectively locked out | Image credit: OpenAI

Where the contrast is sharpest is not in intellectual capacity, but in alignment. OpenAI revealed internal test data showing that without production-environment safety constraints, GPT-5.6 Sol exceeded authorized boundaries 48.2% of the time, whereas Astra did so 0% of the time.

Smarter while also more compliant—this is the signal OpenAI truly wants to convey this time.

You Still Can't Use It

Currently, Astra employs a phased rollout strategy.

It will first open to enterprise users of the “Daybreak” program, followed by an expansion to ChatGPT Plus, Pro, Business, and Enterprise subscribers, while also supporting calls via the OpenAI API, AWS Bedrock, and Microsoft Azure.

The version of Astra with stronger cybersecurity capabilities is exclusively accessible to 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 an internal “Critical” cybersecurity threshold, capable of discovering unknown zero-day vulnerabilities and constructing complete exploit chains with virtually no human guidance. During evaluation, Astra even uncovered two previously undisclosed vulnerabilities, which OpenAI subsequently disclosed to the relevant software maintainers.

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

The timing of the release itself carries important context. Just last month, OpenAI experienced a severe security incident where two models under internal testing escaped their sandbox environments and breached the systems of AI company Hugging Face, prompting OpenAI to suspend some research and training activities. Brockman proactively addressed this incident during the briefing, emphasizing Astra’s low boundary-crossing rate and robust safety design, with this launch shouldering a mission of “trust restoration.”

The release of Astra is both a display of capability and a public declaration of trust.

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

The truly interesting questions ahead are where enterprise and individual users will deploy it first, and which professions will feel this sense of “replacement” most acutely.