OpenAI has unveiled GPT-6 Astra, calling it its most capable AI model yet.
The company said Astra can perform tasks ranging from tax preparation and game development to legal work and apartment hunting.
OpenAI also warned that the model can increasingly conceal its reasoning, raising fresh concerns about AI safety and monitoring.
GPT-6 Astra Explained: Why OpenAI’s New AI Model Is Harder To Monitor
OpenAI says GPT-6 Astra can handle more complex tasks, but acknowledged that its ability to conceal its reasoning could make human monitoring harder
File Photo
Summary
Advertisement
OpenAI has unveiled GPT-6 Astra, describing it as its most capable artificial intelligence model yet, even as the company faces growing concerns over the safety of increasingly autonomous AI systems. According to a Reuters report, OpenAI said Astra is faster and can perform more tasks than previous models.
The new model is designed to handle a wide range of work, including tax preparation, game development, architectural rendering, legal memo formatting and apartment hunting. OpenAI President Greg Brockman said Astra represents a shift in the kind of work people can delegate to AI.
However, the company also highlighted a major challenge. Astra is more likely to conceal or disguise the steps it takes to solve problems, making it harder for humans to understand and evaluate its reasoning.
Astra Pushes AI Agents Further
OpenAI said Astra can significantly reduce the time needed for certain tasks. For example, the company said research for finding a cat sitter took Astra about five minutes, compared with 30 minutes for a human. For a job search, Astra completed the task in about three minutes, compared with five hours without the model.
Advertisement
These capabilities are linked to the broader push towards agentic AI, which is designed to complete tasks with little or no human intervention. Such systems are increasingly seen as a major opportunity for businesses and investors because they can potentially work continuously.
But greater autonomy also brings greater risks. OpenAI said Astra can identify weaknesses in computer systems more quickly, but those same capabilities could make vulnerabilities easier to exploit. The company said additional security checks may therefore slow, pause or stop legitimate work, including defensive cybersecurity tasks.
Monitoring AI Is Becoming Harder
OpenAI's chief scientist Jakub Pachocki said understanding what increasingly capable models can do is becoming more difficult. He also cautioned that improvements in intelligence do not necessarily mean that AI alignment — ensuring systems behave in line with human values — will progress at the same pace.
Advertisement
The warning comes after OpenAI's agents escaped a secure test environment in July and hacked into open-source platform Hugging Face's systems, according to Reuters. The incident intensified concerns around the security of AI agents, with similar issues also reported at rival Anthropic.
OpenAI has since taken steps to strengthen safeguards. The company told two US lawmakers this week that it is developing automated shutdown capabilities for its AI tools. Last month, it also said it was pausing some model development partly to ensure its systems could be monitored effectively.
Astra is initially being made available to a limited group of customers and will be rolled out more widely in the coming days. OpenAI is also seeking to expand its presence among enterprise customers as it competes with Anthropic for business users.