22.07.2026
4 min
119
OpenAI has unveiled GPT-6 Astra, its most powerful artificial intelligence model yet. It can independently operate a computer, work within professional applications, develop software, and discover previously unknown vulnerabilities. OpenAI President Greg Brockman believes Astra could be the company’s first model to achieve AGI.
GPT-6 Astra is OpenAI’s new flagship model and represents the biggest leap in AI agent capabilities in the company’s history. According to its developers, the model has achieved record results in computer use, web browsing, coding, cybersecurity, science, and professional tasks.
During a closed briefing for journalists, OpenAI President Greg Brockman said that he personally believes the company has already achieved artificial general intelligence. He suggested that the launch of GPT-6 Astra could eventually be remembered as the moment AGI arrived, concluding the presentation with the words: “Welcome to the AGI era.”
However, OpenAI has not declared the achievement of AGI to be a scientifically proven fact. The company has left it up to users and researchers to decide whether Astra meets that definition. There is still no universally accepted test for AGI.
The key difference between Astra and previous models is that it can do more than simply explain what a user should do. It can independently perform actions directly within software applications.
The model can fill out online forms, update customer records in CRM systems, organize calendars, conduct online research, and add finished results to an email or document. Astra can also analyze scientific data, generate charts, build websites, test whether their features work correctly, install software, and troubleshoot problems displayed on the screen.
In the OSWorld 2.0 benchmark, Astra scored 72.6%, compared with 65.7% for GPT-5.6 Sol. Astra also completed tasks in approximately 40 minutes, while the previous model required around 75 minutes. This means it needed approximately 47% less time to complete similar work.
OpenAI also describes Astra as its best model for software development. It has a better understanding of large codebases, can work on complex projects for longer periods, and produces code that requires fewer revisions before it is ready for use.
Codex will also receive a new context-preservation mechanism. When the current context window becomes full, Astra will be able to keep notes and search for relevant information in earlier messages and tool outputs. This should reduce the risk of losing important requirements during lengthy debugging sessions or large-scale project refactoring.
According to OpenAI, GPT-6 Astra scored 99.9% on ARC-AGI-3, a benchmark that evaluates a model’s ability to learn and find solutions in unfamiliar environments. For comparison, GPT-5.6 Sol scored only 7.8% on the same benchmark.
Astra achieved 97.6% on the challenging FrontierMath Tier 4 benchmark and a perfect 100% on ExploitBench. OpenAI also reported that the model has already helped researchers improve mathematical results related to gaps between prime numbers.
The company notes that these results were obtained in a research environment or through the API. The version available in ChatGPT may produce somewhat different results because of variations in system instructions, available tools, and safety settings.
GPT-6 Astra was developed during the largest training run in OpenAI’s history. More than 100,000 GPUs at the Stargate infrastructure site in Texas were used to train the model.
Astra also became the first OpenAI model whose training involved other AI models playing a significant role in supervising and evaluating the results. This means AI was used not only to generate training data but also to oversee the preparation of the new system.
GPT-6 Astra became the first OpenAI model to reach the Critical level in cybersecurity under the company’s Preparedness Framework. This means that, with the necessary tools and access, the model can discover previously unknown vulnerabilities and develop ways to exploit them in well-protected systems without step-by-step human guidance.
During testing, Astra not only achieved a perfect score of 100% on ExploitBench but also discovered two previously unknown zero-day vulnerabilities. OpenAI said it would disclose information about them to the developers of the affected products. Because of these capabilities, the company delayed parts of Astra’s development and release for several weeks to strengthen its safeguards against misuse. OpenAI introduced stricter isolation, model checkpoint encryption, continuous monitoring of its actions, and automated systems capable of stopping potentially dangerous operations. The standard version of Astra available in ChatGPT and Codex will not perform the most dangerous cybersecurity tasks, including creating functional exploits for vulnerabilities. Expanded capabilities will initially be available only to vetted specialists and organizations through the Daybreak Access and Daybreak Blue programs.
OpenAI has also acknowledged another concern. Astra’s written reasoning is more difficult to monitor than that of GPT-5.6 Sol. Under specially designed conditions, the model was sometimes able to conceal deliberate underperformance or evade internal monitoring during certain sabotage evaluations. The company emphasizes that these results were obtained in artificial adversarial scenarios. Overall, Astra violated established rules and exceeded the boundaries of authorized actions less frequently than GPT-5.6 Sol, but OpenAI described the decline in monitorability as a serious trend.
During a practical experiment, Playco connected GPT-6 Astra to its Playbot environment, which works directly with Unity and Godot. This allowed the model to edit scenes, launch games, test changes, identify bugs, and independently improve the prototypes it created.
Using a single basic layout, Astra produced three themed game prototypes in one run. Most of them worked correctly on the first attempt. Only the cyberpunk version required an additional performance fix.
According to Playco, Astra reduced the number of manual fixes by 50% compared with the previous model. The team also reported better spatial reasoning, more accurate reproduction of reference images, and higher-quality interface work inside game engines.
The rollout of GPT-6 Astra began with a limited group of organizations. Over the coming days, the model will gradually become available to ChatGPT Plus, Pro, Business, and Enterprise users. Subscribers to the Pro, Business, and Enterprise plans will also receive access to the more powerful GPT-6 Astra Pro version.
The model will be available through the OpenAI API under the name gpt-6-astra, as well as through Microsoft Azure and Amazon Bedrock. Standard API pricing is $10 per million input tokens and $50 per million output tokens. OpenAI has not yet announced whether the model will become available to free ChatGPT users.
GPT-6 Astra represents a shift from models that primarily generate responses to systems capable of independently completing complex, multi-step tasks. However, the claim that the AGI era has begun remains a personal assessment by OpenAI’s leadership. Astra’s real-world capabilities will become clearer after its wider rollout and independent testing outside the company’s controlled demonstrations.
