OpenAI has announced GPT-6 Astra, its latest AI model and the successor to GPT-5.6 Sol. The company describes the newer model as its most intelligent and aligned one yet, positioning it as a major step towards the much-discussed era of artificial general intelligence (AGI).
Astra was trained using more than 100,000 GPUs, making it one of the largest AI training runs to date. OpenAI says the model delivers significant improvements in reasoning, programming, cybersecurity and agent-based tasks, while also being able to complete certain tasks in fewer steps and with fewer tokens.
One of the more notable changes is Astra’s ability to continue working while waiting for an answer from the user. For example, if it needs clarification before completing a particular task, it can ask a question and continue with other related work in the background instead of stopping altogether.
Fewer Hallucinations, Better Performance
OpenAI says Astra also makes progress in factual accuracy, with the model producing substantially fewer factual errors and hallucinations than GPT-5.6 Sol. These improvements are particularly pronounced when Astra operates at lower latency or reasoning settings, suggesting that users may not have to trade speed for accuracy as often.
The model also shows notable gains in computer use and other real-world tasks. OpenAI says Astra can complete computer-based tasks in about 47% less time than GPT-5.6 Sol on its OSWorld 2.0 evaluation, while achieving a higher score of 72.6% compared with 65.7% for its predecessor.
Astra also improves on several cybersecurity benchmarks. These include a perfect 100% score on ExploitBench and a higher 42.4% score on ExploitGym compared with 30.3% for GPT-5.6 Sol.
New Safety Concerns
Improvements aside, Astra also brings new safety concerns. OpenAI has classified it as the first model to reach its “Critical” cybersecurity capability threshold, meaning it can find previously unknown vulnerabilities and develop ways to exploit well-protected systems without human guidance. On the ExploitBench cybersecurity benchmark, OpenAI says Astra achieved a perfect 100% score, prompting the company to restrict its most advanced cybersecurity capabilities to selected testing partners and introduce additional safeguards.
Despite this, OpenAI says Astra is generally safer than its predecessors in higher-risk scenarios, with improvements in areas such as emotional reliance, self-harm requests and interactions with users under 18. The company has also strengthened its protections against malicious cyber use and unauthorised actions, although it acknowledges that these safeguards could occasionally interfere with legitimate tasks.
The bigger concern is that Astra may be harder to monitor than GPT-5.6 Sol. OpenAI says the model has greater control over its chain-of-thought and can sometimes avoid detection during adversarial evaluations, including when deliberately underperforming or attempting certain sabotage tasks. This reduced monitorability could make it more difficult to audit the model as its capabilities become increasingly autonomous.
Availability And Pricing
GPT-6 Astra is beginning its rollout today, but it will not immediately be available to everyone. OpenAI says select organisations will receive access at launch, while ChatGPT Plus, Pro, Business and Enterprise users are expected to gain access over the coming days.
The model will also be available through the OpenAI API and AWS. API pricing starts at US$10 per million input tokens and US$50 per million output tokens.
