OpenAI's strongest AI causes mixed reactions

"From ChatGPT to Astra took place in four years. AGI has appeared. Congratulations to the OpenAI team," Nvidia CEO Jensen Huang wrote on Twitter on September 6.

OpenAI launched GPT-6 Astra on September 3, calling it "the world's most intelligent and relevant model", capable of performing "rigorous work with outstanding speed, accuracy and judgment", "the new standard for computer use, web browsing, software engineering, cybersecurity, science and professional work". On the blog, OpenAI President Greg Brockman also identified the product as an AGI - defined as an AI system with the ability to think equal to or surpass humans.

OpenAI defines GPT-6 Astra as "highly automated systems that outperform humans in most jobs of high economic value", a major breakthrough in research, a change in "what is possible" that humans entrust to AI. The task-focused model requires reasoning, programming, and automated work on a computer that can handle multi-step processes, using browsers and software to perform tasks ranging from research and data analysis to creating documents, spreadsheets, and presentations.

According to OpenAI, Astra scored 72.6% on the OSWorld 2.0 test - a tool that evaluates an AI agent's ability to control a computer in carrying out complex and lengthy workflows; and completed tasks about 47% faster than GPT-5.6 Sol in latency simulation. It is also the first model to reach a critical threshold in cybersecurity, meaning it can find unknown software vulnerabilities and build attacks that work without humans. Sam Altman's company said that GPT-6 Astra scored 100% on ExploitBench - a tool to evaluate the actor's offensive cybersecurity capabilities and LLM. In a test by the company, the new AI also discovered two previously unknown zero-day vulnerabilities.

"Welcome to the era of AGI superintelligence," Brockman wrote. "Personally, I think OpenAI has achieved its goal."

According to Decrypt, after only 48 hours of launch, GPT-6 Astra received positive feedback, especially for its ability to work on "anything related to space, mechanics or agents". Investor Matt Shumer, former CEO of HyperWrite, tried using Astra combined with the game development tool Unreal Engine to successfully create a detailed 3D map of the Manhattan area with "every street is perfect". He also asked Astra to build a survival world with characters with unique personalities, then let it run overnight. A day later, he heard voices from the office and discovered that the characters "started talking to each other".

Independent developer Tom Krcha also gave Astra a single image of the house, then ordered a complete reconstruction, including the interior. The video created a short time later surprised Krcha. "Now everyone in the world has a 3D designer at their fingertips," he wrote on Twitter.

A Chinese developer nicknamed SuSu has implemented a similar idea at a city scale. This person ordered Astra to recreate the city of Hangzhou and surrounding towns using Three.js - a java script library that allows displaying 3D graphics right in the web browser without downloading. Within 24 minutes, the tool creates interactive maps with West Lake, Lieu Phong Pagoda, terraced fields or wetlands... with day and night display options.

Gaming is the area where the GPT-6 Astra excels, according to Creative Bloq. For example, game company Playco said it tried creating a game using this AI and found "only minimal human intervention", which reduced the number of manual fixes during the prototyping process by 50%.

Anshu Chimala, a UX/UI designer and AI developer at Apple, tried creating a 3D game in 45 minutes and "got it done in just one try". He describes a process that only takes up a few percent of the allowed memory capacity but still produces an isometric graphic style product (a method of depicting images of 3D objects in two-dimensional space) with good characters, environments and overall aesthetics. "Astra is a superfast AGI for 3D games," Chimala wrote on Reddit.

The GPT-6 Astra running agent can also automate almost the entire gaming process. Twitter account cozyblaze uses this AI to play Portal - Valve's classic first-person puzzle game. The AI ​​completed the entire game in less than 24 hours at a token cost of 571.18 USD. According to The Verge, this level is "equivalent to human level".

However, according to Creative Bloq, despite being a breakthrough that makes the 3D modeling and game development process faster and more accessible, many people worry that GPT-6 Astra can be exploited as a copyright infringement tool, allowing people to copy entire game ideas and sell them as their own products. Others say that while the demos look impressive, Astra doesn't outperform Anthropic's Claude Fable in performance tests.

Besides, GPT-6 Astra is not highly appreciated for its creativity, especially without strong guidance, context and direction. Developer Louis-François Bouchard used an internal evaluation system to score this AI's writing ability and found that the model ranked 11th with 1,995 points, lower than the previous GPT-5.6 Sol with 2,156 points, less than the leading position Claude Fable 5 with 2,312 points. Astra costs about 0.26 USD per scenario, about 1.8 times more than Sol.

According to the Artificial Intelligence Index (AII), which synthesizes and evaluates AI's reasoning, knowledge and programming abilities, Astra scored 61.2 points. This number is lower than the 60.9 points of GPT-5.6 Sol and 65.7 points of Claude Fable 5.1 but the price is 2.5 times the price of Sol.

New York-based independent researcher Giuseppe Paleologo tested and rated the "creativity-related" results of the GPT-6 Astra as "a mixture of the obvious and the exaggerated." According to him, true creativity is still very far away in this model.

Mia AI Lab, a group on Hugging Face that specializes in providing the latest open source AI/LLM configurations, optimization formats, or clones, has a similar view. According to the group, the GPT-6 Astra may be the best model for some tasks, but is "boring and lacking in personality", and advises "it should be avoided for any creative work".

Mr. Gary Marcus, a famous AI researcher and critic in Silicon Valley, believes that OpenAI's latest model is not enough to reach the AGI level. Based on his self-drafted definition of 10 differences between AGI and conventional AI, he said GPT-6 Astra only meets one or two criteria. "Astra still has many shortcomings," Marcus told Business Insider.

However, thinking ability is the most worrying point. According to The Information, GPT-6 Astra uses a reasoning technique called "recurrent depth" that allows it to operate beyond the sequential thinking typical of most AI inference. This technique, also known as "opaque recurrence", has the potential to make the model's train of thought more difficult to monitor.

According to TechCrunch, reasoning AI systems currently use chains of thought (CoT), displayed right on tools to help humans see through the "steps" of the model, thereby helping to detect deviant or out-of-control behavior. However, the "repeat depth" technique on GPT-6 Astra processes a query multiple times in a closed loop, making monitoring increasingly difficult by leaving very little readable trace. Accordingly, it almost ignores the traditional thought chain mechanism.

"Astra uses an unclear iterative method, which is truly worrying," Buck Shlegeris, CEO of the non-profit research organization specializing in AI safety and security Redwood Research (USA), wrote on Twitter. "If OpenAI continues to develop this technique, they will completely destroy the ability to monitor CoT."

Zvi Mowshowitz, an AI safety campaigner in the US, said OpenAI is "playing with fire" when using this technique for the new model. According to him, the use of "repeat depth" also violates the taboo that the company and Anthropic themselves fought to establish: the effort to maintain integrity and surveillance for as long as possible.

Ryan Greenblatt, chief scientist at Redwood Research, also believes that OpenAI's approach makes future models more difficult to predict due to "inference in the implicit space". "I hope it's not too late and OpenAI will stop here," Greenblatt wrote on the blog.

OpenAI has not yet commented. On Twitter, OpenAI chief scientist Jakub Pachocki said the company "works to preserve and use supervised inference chains since the first models," but did not mention the new inference engine on GPT-6 Astra.