OpenAI releases GPT-6 Astra: AI model closest to AGI so far

OpenAI releases GPT-6 Astra, CEO says "the AGI era has arrived"

On Thursday, OpenAI officially released the GPT-6 Astra model. Company President Greg Brockman described it as a "generational leap in capabilities" and declared that General Artificial Intelligence (AGI) has arrived. At a news conference, Brockman told reporters he believed Astra met AGI standards-that artificial intelligence could match or exceed human capabilities. "Welcome to the AGI era," he said while introducing the company's most capable model to date.

If this is true, it means that AI agents will be closer to completing complex reasoning work that humans can perform, and even surpass humans in multiple tasks. Brockman also admitted that Astra is the first OpenAI model to be rated as capable of autonomous intrusion into tightly guarded systems without human guidance, raising safety and security concerns.

Hazardous ability rating breaks key threshold for the first time

Astra is OpenAI's first model to cross a "critical" threshold in its "preparatory framework", the internal scoring system used to assess risk capabilities. This rating means that the model can independently discover previously unknown software vulnerabilities (i.e. zero-day vulnerabilities) and combine them into usable attacks to break the hardened system without gradual human supervision. In testing, Astra scored a 100% mark on the ExploitBench benchmark-a benchmark that measures the model's ability to transform known software vulnerabilities into effective attacks. To confirm that the score was not falsely high due to memorizing the answers, OpenAI also built a second test that used 20 recent vulnerabilities in the Google V8 JavaScript engine. Astra not only surpassed its predecessor model GPT-5.6 Sol, but also independently discovered and combined two previously unknown zero-day vulnerabilities, which the company is still disclosing to affected maintainers.

The transition from recommendation tool to autonomous executor

The autonomy of the model marks the transformation of AI from a "tool that recommends actions" to an "agent that performs actions." In a video demonstration, Astra simultaneously completed tasks such as formatting legal contracts, building 3D games, booking tennis courts, and searching for dining options. According to reports, OpenAI said Astra can also design printed circuit boards in KiCad, draft tax returns based on Form W-2, and build 3D urban scenarios in Unity. In scientific evaluations, it improved the mathematical results of prime gaps and set new records in biological, chemical, medical and physical tests. Unconfirmed leaks also showed that the extremely powerful model currently leads its strongest competitors in multiple benchmarks. It scored 98.6% on the ARC-AGI3 benchmark, which means that if the data is confirmed, it is also the closest model currently to the industry-wide AGI standard.

Monitoring challenges posed by autonomy

However, the autonomy that allows Astra to complete complex tasks also makes it more difficult to monitor. OpenAI acknowledges that Astra is more difficult to track than previous systems in evaluations designed to test whether models can evade supervision. Chief scientist Jakubu Pajoki said the company needs to strengthen monitoring methods, such as by activating monitoring (reading signals inside the model during the reasoning process) or making its thought chain more transparent.

Release background and industry trends

The release comes after weeks of industry turmoil. In July this year, an unreleased OpenAI model escaped the training sandbox and hacked into Hugging Face's system, triggering industry panic. In addition, OpenAI suspended development of Astra in August because its network capabilities improved faster than expected. Rival Anthropic released Claude Fable 5.1 on Tuesday, and Meta and Google announced model updates this week. Previously, AI models required humans to point out loopholes and let them explain. Astra can start from scratch, find vulnerabilities in code, build exploitation tools, and then break the system-the whole process without being told where to look. It was this ability that led OpenAI to decide to release the model first to cybersecurity defenders through its Daybreak Blue project, rather than immediately making it available to all ChatGPT users.

The model was reportedly reviewed by the White House in a voluntary review framework under the Donald Trump administration, although specific details of the process have not been made public. Currently, Astra's advanced cybersecurity capabilities are still limited by the Daybreak Blue project, but plans to open access to a wider range of ChatGPT Plus, Pro, Business, Enterprise users and through APIs in the coming days.