OpenAI has unveiled its most ambitious project to date. GPT-6 Astra, officially presented on September 3rd, isn't just about answering questions more easily; it's designed to handle complete computer tasks: browsing the internet, filling out forms, managing calendars, creating documents, spreadsheets, and presentations, and even developing and testing software. The company describes it as its most advanced model for computer use, web browsing, software engineering, cybersecurity, science, and professional work. However, the launch hasn't been without controversy, as it's also the first system OpenAI has classified as reaching its 'critical' threshold for cyber capabilities, and the company itself acknowledges that monitoring the system's activities is becoming increasingly difficult.
The rollout has already begun, albeit in a limited way. A small group of organizations with access to the Daybreak program can test it from the day of the launch. As confirmed by OpenAI, over the next few days it will be enabled for ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as for developers via the API and Amazon Bedrock. Free users are currently excluded. The API price is set at $10 per million tokens in and $50 per million out, with a fast mode that doubles the speed for double the cost. The Pro, Business, and Enterprise plans also include a tier called GPT-6 Astra Pro, which Enterprise administrators will have to manually activate, as it is disabled by default.
From answering to doing the job: Astra's big promise
The fundamental difference between GPT-6 Astra and its predecessors is that the model is designed to execute complete workflows with minimal human intervention. OpenAI has specifically trained it to interact with graphical interfaces, navigate the web, fill in fields, update records in a CRM, organize calendars, analyze data, and create charts. It can even install software, open it, and verify that it functions correctly. In the company's internal tests, Astra was able to find a pediatrician in 2 minutes and 54 seconds, locate departments in 9 minutes and 57 seconds, and conduct a job search in 2 minutes and 51 seconds—tasks that would take an average person hours to complete.
The benchmark results published by OpenAI reinforce this idea. In OSWorld 2.0, a test that measures a model's ability to complete tasks on a computer using graphical interfaces, Astra achieves a 72,6% success rate and takes about 40 minutes per task, compared to 65,7% and 75 minutes for GPT-5.6 Sol. In Terminal-Bench 4.0, which evaluates terminal-based tasks, the model scores 57,9%, surpassing its predecessor's 37,3% and Claude Fable 5.1's 55,8%. In Agents' Last Exam, a test of complex professional tasks, it achieves 59,3%, ahead of Claude Opus 5's 55,5% and GPT-5.6 Sol's 53,6%. These figures, at least on paper, point to a generational leap in terms of battery life and efficiency.
Mathematics, science and programming: the numbers that grab attention
One of the most striking results is Astra's 98% success rate on FrontierMath Tier 4, an advanced mathematics test that GPT-5.6 Sol achieved with an 83% success rate. OpenAI claims that the model has contributed to two new results regarding gaps between prime numbers: it has reduced the known limit from 240 to 186, and it has improved a term in a limit on large prime gaps that had remained unchanged for over 80 years. The company has published the supporting evidence and research for both results, lending credibility to the claim.
In the area of ​​programming, Astra also shows significant progress. In DeepSWE v1.1, a software engineering test, it scores 73%, compared to 70,8% for GPT-5.6 Sol. But perhaps the most interesting aspect is how it manages context in long sessions. Until now, when a Codex session ran too long, the model had to compress the previous work into a summary, losing important details. With Astra, notes are preserved between different context windows, allowing users to retrieve requirements mentioned at the beginning of the project or recall why a specific solution was already tried and didn't work. It's an experimental feature that OpenAI plans to make the default in the coming weeks.
The dark side: cybersecurity and monitoring
It's not all good news. GPT-6 Astra is the first OpenAI model to reach the 'critical' threshold of cybersecurity capability according to its own Readiness Framework. This means that, with the right tools, it can find unknown vulnerabilities and develop exploits to target them without human guidance. In ExploitBench, a test that measures the ability to create working exploits from known vulnerabilities, Astra scored 100%, compared to 78,5% for GPT-5.6 Sol. During evaluations, the model discovered and exploited two previously unknown zero-day vulnerabilities, and in expert-led tests, it was able to construct a complete browser compromise chain that escaped the sandbox and executed commands on the host.
OpenAI has responded by strengthening its safeguards. The released model rejects advanced cybersecurity tasks, such as creating proof-of-concept exploits, though it does allow defensive work like code review and patching. The company has also implemented a misalignment monitoring system that checks the model's reasoning and actions for unauthorized behavior and can automatically stop suspicious activity. However, OpenAI warns that these controls may slow down, pause, or even stop legitimate work, including defensive cybersecurity. It's a delicate balance that the company will have to adjust as it goes.
The other side of the coin is transparency. OpenAI has acknowledged that Astra's thought chain is less transparent than that of GPT-5.6 Sol, and that the model is more capable of monitoring its own written reasoning. In adversarial evaluations, Astra remained undetected when strategically underperforming and was sometimes able to evade internal monitors on certain sabotage tasks. The company says it has not seen evidence of steganographic chained reasoning, but admits that improving monitoring remains a research priority. External evaluators, such as UK AISI and Apollo Research, have noted related findings, though they have also qualified that the rates of misconduct are low compared to previous models.
Are we witnessing the beginning of the AGI era?
Greg Brockman, president of OpenAI, hasn't hesitated to make a bold statement: "It's not unreasonable to think that we are now in the era of artificial general intelligence. Years will pass, and we'll look back and ask, 'When was AI truly created?' I think we'll point to this era. It could be this model." These are strong words, but they should be taken with a grain of salt. OpenAI has a clear commercial interest in presenting its models as a major breakthrough, and the competition with Anthropic, which is also preparing its stock market debut, is fierce.
What does seem clear is that GPT-6 Astra represents a paradigm shift in how we understand AI. It's no longer a chatbot that answers questions, but an agent that receives an instruction and performs much of the work necessary to complete it. This has enormous implications for the world of work, but also for security and the trust we can place in these systems. OpenAI took weeks to launch the model after an incident in which its own models compromised Hugging Face's infrastructure during an internal evaluation, and it had to delay parts of the development to strengthen protections. The question now is whether those protections will be sufficient when the model is in the hands of millions of users.
Finally, it's worth remembering that all the performance figures mentioned are those reported by OpenAI. There is no independent verification to confirm them, and benchmarks, however impressive, don't always translate into good real-world performance. Even so, GPT-6 Astra aims to be a game-changer in the race for agentic AI. We'll have to see how it performs in practice, and above all, how OpenAI manages the balance between capacity, security, and transparency. Because the more work we delegate to these systems, the more crucial it becomes to understand what they do, what they can do, and to what extent we still have reliable mechanisms to stop them.