Anthropic AI development speed doubles despite CEO calls for caution

Accelerated development cycles

Anthropic is currently developing its artificial intelligence systems at a pace that has doubled over the last year. This finding comes from an analysis by Model Evaluation and Threat Research (METR), an independent nonprofit based in Berkeley, California. The research group was granted unprecedented access to Anthropic's internal data and staff interviews to study these trends.

The report suggests that the company is making significant strides in improving models to perform well-specified tasks. This rapid growth is particularly notable because it occurs while the company prepares for a major initial public offering (IPO) later this month. The IPO is expected to be one of the largest initial public offerings of stock ever held in the industry.

While the company is moving quickly, the data shows that there are currently no major bottlenecks like data shortages or computing power limits. The research indicates that the company is not being slowed down by the physical resources typically required to train large models. Instead, the focus has shifted toward the efficiency of the development process itself.

The role of recursive AI

A primary reason for this increased speed appears to be the use of AI to improve other AI. Anthropic has noted that its Claude chatbot is increasingly helping the company build newer and more advanced versions of its own technology. This process is often referred to as recursive self-improvement, where a system's ability to build the next generation of AI becomes a core driver of progress.

METR researchers ruled out the growth of computing power as the main reason for the acceleration. In traditional AI development, the amount of compute—the processing power used to train models—is usually the direct link to how well a model performs. Because compute growth isn't the primary driver here, the focus turns to the software and methodology used to train the systems.

The report notes that while AI is not yet making fully autonomous contributions to the most important parts of training, there are signs of life suggesting this could happen soon. This creates a potential scenario where the acceleration could begin to compound. As the models get better at helping humans build better models, the speed of improvement may increase exponentially.

Safety vs. speed tensions

The rapid pace of development stands in contrast to recent public statements from Anthropic's leadership. CEO Dario Amodei recently urged the industry to slow down the pace of improvement so that safety measures can catch up. He specifically called for a slower approach to ensure that the risks associated with advanced models do not outpace our ability to manage them.

Despite these calls for caution, the internal data shows the company is still moving at a high velocity. Anthropic's co-founder Sam McCandlish acknowledged that while AI is speeding up research, it is difficult to measure the exact extent of that impact. He noted that the company decided not to share the specific details of what is driving the speedup because it is commercially sensitive.

McCandlish also argued that sharing the exact roadmap for these speedups could provide a blueprint for other companies to move even faster. He emphasized that the current pace of progress has significant safety implications because it determines how quickly the risks from advanced models increase. The company maintains that external evaluators are necessary to help pace this frontier development.

Transparency and oversight

The METR report was based on an 11-week assessment involving internal data and employee interviews. During this time, the nonprofit reviewed both public and private risk reports provided by Anthropic. The researchers found that the company's public report actually understated the level of acceleration that was occurring internally.

Anthropic's redactions in the report limited how much detail the nonprofit could share with the public. The company argues that protecting its proprietary methods is necessary for competitive reasons. However, the lack of public detail makes it difficult for external observers and policymakers to get a clear sense of the actual rate of progress.

The involvement of METR is part of a broader conversation about how AI companies should police themselves. The Trump administration has encouraged companies to manage their own safety protocols. By inviting independent researchers to look at internal data, Anthropic aims to demonstrate a commitment to safety while still pursuing rapid technological growth.