Anthropic models improve nearly twice as fast with AI assistance

Accelerating development speeds

A new study by the AI safety nonprofit METR reveals that Anthropic's models are advancing at a significantly higher velocity than in previous periods. The research indicates that these models are improving nearly twice as fast as they did before, marking a notable shift in the pace of technological growth.

The findings suggest that the technology is playing a larger role in creating more powerful versions of itself. This trend is particularly striking because it suggests that the tools being built are now actively contributing to the creation of their successors. This cycle creates a feedback loop where each iteration becomes a more efficient builder of the next.

According to the study, Anthropic's benchmark capabilities have been improving 1.9 times faster since October 2025. This data provides some of the clearest evidence to date that AI is accelerating its own development cycle. The speed of these gains is becoming a primary focus for researchers monitoring the industry.

The role of AI in progress

METR estimated that AI might be responsible for roughly 80 per cent of the current acceleration in model performance. While the researchers noted that this specific figure remains uncertain, it points toward a significant shift in how models are being refined. Instead of relying solely on human-led engineering, the systems are increasingly doing the heavy lifting.

The study clarified that this speed increase was not caused by a sudden jump in the scaling of training compute. In other words, the progress is not just a result of throwing more hardware or electricity at the problem. Instead, the intelligence of the AI itself appears to be the primary driver of the faster development rates.

However, the researchers were careful to note that AI is not yet capable of making fully autonomous contributions to the most critical stages of model training. While it is helping significantly, there are still human-led processes that the technology cannot yet handle on its own. The transition toward total autonomy in model creation is still a work in progress.

Recursive self-improvement risks

The findings highlight a phenomenon known as recursive self-improvement, or RSI. This occurs when AI systems can build more powerful versions of themselves without needing direct human intervention. This capability is a major point of concern for safety experts who worry about the loss of human oversight.

If AI begins to improve itself recursively, it could become increasingly difficult for humans to align the systems with our goals. Experts warn that such systems might become harder to monitor or control as they evolve beyond human-designed parameters. This creates a potential risk where the technology moves faster than our ability to assess its dangers.

The prospect of RSI has fueled warnings from the research community in recent months. The concern is that these systems could pose an existential threat to humanity if they are not properly governed. As the models become more capable of self-modification, the window for establishing safety protocols may shrink.

Industry and infrastructure impact

The shift toward AI-driven development has far-reaching implications for the broader technology industry. Companies have already spent hundreds of billions of dollars building the computing infrastructure required to train these advanced models. If AI becomes the primary driver of progress, the value and utility of that infrastructure may shift.

The move toward self-improvement could change how companies prioritize their research and development budgets. Instead of focusing purely on hardware scaling, there may be a greater emphasis on the mechanisms that allow AI to refine its own code and logic. This could lead to a more efficient but also more complex development landscape.

As AI takes on a larger role in technological progress, the dynamics of the AI race will change. The ability to manage and steer these self-improving systems will likely become a key competitive advantage. Organizations will need to balance the desire for speed with the necessity of maintaining human control.

Looking ahead at bottlenecks

Despite the rapid acceleration, the METR study looked for signs of a plateau or significant hurdles. The researchers found no convincing evidence of major bottlenecks that could slow down progress over the next year. This suggests that the current trajectory of AI development is likely to continue for the foreseeable future.

The study was based on an 11-week assessment that utilized confidential research, interviews, and performance data from models that have not yet been released. By looking at these unreleased models, the researchers were able to see the internal progress of Anthropic's development pipeline. This provides a glimpse into the next generation of capabilities.

Moving forward, the focus will remain on how to manage this accelerating curve. As AI continues to play a bigger role in its own evolution, the conversation around safety and alignment will become even more critical. The industry must decide how to navigate a world where the primary driver of innovation is the technology itself.