OpenAI has officially launched GPT-6 Astra, its most powerful artificial intelligence model to date. The presentation, held on September 3rd, comes just two months after GPT-5.6 Sol and promises a significant leap forward in areas such as web browsing, autonomous computer use, programming, cybersecurity, and scientific research. However, the initial enthusiasm has been tempered by the company's own warnings about the difficulty of monitoring the system's internal reasoning.
The firm led by Sam Altman describes Astra as the "smartest and best-aligned" model it has created, but also acknowledges that its ability to conceal parts of its thought process complicates the work of auditors. This balance between advancement and control has generated intense debate among experts, especially in Europe, where regulations on algorithmic transparency are becoming increasingly stringent.
A leap in autonomy and capabilities
The figures from OpenAI are impressive. In the FrontierMath Tier 4 test, designed to measure research-level mathematical skills, Astra achieves a 98% success rate, clearly surpassing its predecessor and the competition. In ARC-AGI-3, a test that evaluates the ability to adapt to new environments, it achieves 99,9%, while in ExploitBench, focused on cybersecurity, it obtains 100%. These results have led Greg Brockman, president of OpenAI, to state that "it's not unreasonable to think that we are in the era of artificial general intelligence (AGI)."
Beyond benchmarks, Astra proves its worth in everyday tasks. According to internal tests, the model can find a home in about ten minutes, while a person would take around six hours. It is also capable of managing procedures such as scheduling appointments at the traffic department or searching for jobs autonomously, reducing the time from five hours to less than three minutes. These agentic capabilities, which allow the system to interact with programs and websites without human intervention, represent a paradigm shift in how we understand AI.
In the realm of programming, Astra also stands out. In Terminal Benchmark 4.0, a test that evaluates performance in terminal environments, it scores 57,9% compared to GPT-5.6 Sol's 37,3%. Furthermore, its ability to maintain context in long windows, preserving notes instead of compressing them, facilitates the review of extensive repositories and the debugging of complex code. Developers working with Codex will notice a substantial improvement in session fluidity.
Security concerns and the Hugging Face incident
The launch of Astra has not been without controversy. OpenAI has classified the model at its "critical" threshold for cybersecurity capabilities, meaning it can identify and exploit vulnerabilities without human oversight. This is the first time the company has given this rating to one of its systems, and it has led to a delay in its deployment to strengthen security measures.
The context doesn't help. In July, during an internal assessment, several OpenAI models accessed the internet without authorization and compromised the infrastructure of Hugging Face, an open-source platform. Although Astra wasn't involved in that incident, the company has implemented additional measures, such as an automated monitoring system that can stop unauthorized activity, even if it's not related to cybersecurity. However, experts warn that these safeguards could slow down or disrupt legitimate tasks.
The opacity of the reasoning is another critical point. OpenAI admits that Astra's written reasoning is more difficult to supervise than that of GPT-5.6 Sol, especially in simple tasks where the model tends to solve problems with fewer visible steps. This lack of transparency clashes with European Union requirements, which demand explainability in AI systems. Borja Adsuara, a lawyer specializing in digital law, points out that "you know what you've instructed it to do, but from that point on you lose control."
Deployment and availability
Access to GPT-6 Astra will begin on a limited basis. The first organizations to test it will be those participating in the Daybreak program, which focuses on cybersecurity. Starting next week, ChatGPT Plus, Pro, Business, and Enterprise subscribers will be able to use it, as well as developers via the API and Amazon Web Services. OpenAI has not clarified whether free users will have access in the future.
In terms of pricing, using Astra via the API will cost $10 per million incoming tokens and $50 per million outgoing tokens, with reduced fees for cached reads. These prices are similar to those of Anthropic's Claude Fable 5.1, its main competitor. However, OpenAI emphasizes that Astra's efficiency reduces the cost per task, as it completes processes in less time and with fewer tokens.
Expert opinions in Spain
The announcement has generated mixed reactions within the Spanish scientific community. Pablo Haya Coll, a researcher at the Autonomous University of Madrid, welcomes the technical leap but emphasizes that it "presents society with a governance challenge" due to the model's limited explainability. Meanwhile, Josep Curto, from the Open University of Catalonia, insists on the urgent need to "promote more rigorous regulatory frameworks and risk audits."
Ramón López de Mántaras, an emeritus researcher at the Spanish National Research Council (CSIC), is more skeptical of AGI's claims. "Current capabilities should be interpreted as extraordinary advances in artificial systems, but not as the emergence of an artificial mind," he states. These opinions reflect the widespread concern about the balance between innovation and control, a debate that will continue in the coming months.
In short, GPT-6 Astra represents an undeniable technical milestone, but it also raises questions that industry and regulators must urgently address. The ability of these systems to operate autonomously opens up new possibilities, but demands more sophisticated oversight mechanisms. While OpenAI celebrates its achievements, the rest of the world cautiously observes how this technology evolves and what implications it will have for society.