Anthropic AI development speed has doubled in one year

Accelerated growth rates

An independent monitoring group has revealed that Anthropic is developing its artificial intelligence systems around twice as quickly as it did a year ago. The research group, Model Evaluation and Threat Research (METR), conducted a study by reviewing internal data and interviewing staff at the company. This analysis comes as Anthropic prepares for a major public stock offering later this month.

The findings suggest that the company has made as much progress on benchmarks in the last year as it did between October 2023 and October 2025. METR noted that while they provided the company with unprecedented access to internal data, some redactions limited the ability to share more detailed estimates. Despite these limits, the group remained confident in the finding that the growth rate has significantly accelerated.

The report highlights a tension between the company's rapid technical progress and the public statements made by its leadership. While the technology is advancing at a high velocity, Anthropic's CEO has expressed concerns about the pace of these improvements. The company is currently navigating a period of intense scrutiny regarding how quickly it can safely deploy these increasingly capable models.

The role of recursive AI

Researchers believe that the primary driver behind this rapid acceleration is the use of AI to improve other AI. Anthropic has stated that its chatbot, Claude, is increasingly being used to help the company develop newer and more advanced versions of itself. This process is often referred to as recursive self-improvement, where the output of one model becomes the input for the next generation.

METR was able to rule out the growth of computing power as the main reason for this spike in performance. Traditionally, the amount of compute—the processing power used to train models—is the main factor that determines how well an AI performs. Because the acceleration is happening without a corresponding surge in compute, the focus shifts toward the efficiency of the self-improvement cycle.

The report notes that while AI is not yet making fully autonomous contributions to the most important parts of the training process, there are significant signs of life. These signs suggest that the ability for AI to contribute to its own development could happen soon. This creates a potential feedback loop where the speed of improvement could compound, making it harder for humans to predict the trajectory of the technology.

Safety and oversight concerns

The rapid pace of development has sparked concerns that AI models could eventually elude human control with catastrophic consequences. Anthropic CEO Dario Amodei has acknowledged these risks, urging the industry to slow down so that safety measures can catch up. He has specifically highlighted the danger of AI's growing ability to build the next generation of AI without sufficient oversight.

Amodei has championed the idea of having external evaluators embedded within leading AI labs to provide independent checks on progress. He specifically identified METR as an appropriate third-party evaluator for this purpose. This move toward independent oversight is supported by other industry leaders, including the CEO of OpenAI, who also advocates for external evaluators with employee-like access.

The findings are particularly notable because the details on these risks have primarily been provided by the AI companies themselves. This comes at a time when the administration has encouraged these companies to police their own safety standards. The lack of clear bottlenecks, such as data limits or hardware shortages, suggests that the only remaining constraints on progress may be the choices made by the developers.

Limitations of the study

While the findings are significant, the METR report includes several caveats regarding the scope of the acceleration. The researchers noted that the doubling in speed might be limited to well-specified tasks that can be graded automatically. These are tasks with clear instructions, which differ from more messy or complex tasks that are harder to define or grade.

There is also the possibility that Anthropic may have understated its progress in previous years. If earlier milestones were achieved more quickly than reported, the current rate of acceleration might appear larger than it actually is. The researchers acknowledged these internal uncertainties and potential overstatements while maintaining their core conclusion about the growth trend.

Despite these caveats, the study provides a rare look into the internal mechanics of a leading AI lab. It confirms that the technological trajectory is shifting toward a model where AI assists in its own creation. This shift moves the conversation from how much data or power we can provide to how effectively the models can refine their own capabilities.