Garp Independent AI & technology journalism
Sunday, November 22, 2026 Sign In · Join Subscribe
Latest Figure leads vibe-coding movement, transforming gadget customization with AI

AI news, research, models, robotics, chips, startups, and infrastructure coverage.

Updated daily

Home  /  AI News  /  METR warns AI models like Opus 5.5 and GPT-6 Astra are advancing too quickly for safety testing

AI News

METR warns AI models like Opus 5.5 and GPT-6 Astra are advancing too quickly for safety testing

On November 22, 2026, the AI research organization METR reported that leading models such as Opus 5.5 and GPT-6 Astra are rapidly improving in research tasks, raising urgent concerns about their safety and reliability.

According to METR, the current public data is inadequate to rule out a dangerous acceleration in AI research, potentially leaving society ill-prepared to understand and mitigate catastrophic risks posed by these advanced systems. The organization cautioned that while models like Opus 5.5 and GPT-6 Astra excel in well-defined coding and computational tasks, they struggle with open-ended questions that require nuanced judgment and real-world context.

The report highlights that these AI systems frequently outperform experienced human professionals in specific tasks, although they have not yet surpassed large groups of top experts. Remarkably, METR found that the ability of these models to handle longer tasks is increasing at an extraordinary rate, doubling every 3.6 months. If this trend continues, the organization estimates that by July 2027, these systems could achieve a 50 percent success rate on tasks lasting up to 200 hours.

This rapid improvement raises alarms about the potential for AI systems to operate beyond the scope of human oversight, especially as they become more integrated into critical sectors such as healthcare, finance, and transportation. Despite these advancements, METR warns that common tests used to evaluate AI performance may not accurately reflect real-world capabilities. Developers often train models in controlled environments that do not account for the complexities of actual use cases, leading to potential overestimations of their effectiveness.

As AI systems become more prevalent, the implications of these findings raise critical questions about safety and oversight in AI development, emphasizing the need for more rigorous testing and evaluation methods to ensure these technologies are safe for public use.

Why it matters: The rapid advancement of AI models like Opus 5.5 and GPT-6 Astra poses significant risks to safety and reliability, as highlighted by METR's findings. With the potential for these systems to operate beyond human oversight, it is crucial for regulatory bodies to establish more stringent testing protocols to assess their real-world capabilities.

Original reporting: The Decoder