Anthropic AI development speed doubles as models begin self-improvement

Accelerated progress at Anthropic

Anthropic is currently developing its artificial intelligence systems at roughly double the speed it maintained a year ago. 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 assess the company's development pace.

The findings are particularly notable because they come from an external audit rather than a self-reported figure from the company. METR's assessment suggests that Anthropic has made as much progress on benchmarks in the last year as it did between October 2023 and October 2025. This rapid growth is occurring just as the company prepares for its upcoming debut on the public stock market.

The report highlights a significant discrepancy between public and private communications. METR noted that a public risk report released by Anthropic in August actually understated the level of acceleration taking place. This discrepancy underscores the difficulty of monitoring the internal progress of private AI labs that are encouraged to police their own safety standards.

The role of recursive self-improvement

Researchers believe the primary reason for this doubling in speed is the use of AI to improve other AI. Anthropic has confirmed that its Claude chatbot is increasingly being used to help the company develop newer, more advanced versions of the technology. This process of AI building the next generation of AI is often referred to as recursive self-improvement.

METR was able to rule out the growth of computing power as the main driver of this acceleration. Traditionally, the amount of compute used in training is the primary factor in how well a model performs. Because there are no current bottlenecks in data or hardware, the speedup appears to be a result of more efficient, AI-driven research methods.

While the technology is not yet capable of making fully autonomous contributions to the most critical parts of model training, researchers have noted significant signs of life in that direction. The concern is that if AI begins to improve itself, the acceleration could compound, making it harder for humans to predict or intervene in the development cycle.

Safety concerns and industry calls

The rapid pace of development has sparked a debate over whether safety measures can keep up with technological gains. Anthropic CEO Dario Amodei has recently urged the industry to slow down its work. He argued that the pace of improvement must be managed so that safety protocols have enough time to be properly established and tested.

Amodei expressed concern that unchecked recursive self-improvement could outrun the human ability to understand or control these systems. He has championed the idea of having third-party evaluators, like METR, embedded within leading AI labs. This approach aims to provide a clearer picture of internal progress to policymakers who currently lack a strong sense of the actual acceleration rates.

Despite these calls for caution, the company faces a difficult tradeoff regarding transparency. Anthropic has decided not to share the exact details of what is driving its speedup because the information is commercially sensitive. Furthermore, the company fears that providing a specific roadmap for acceleration could be used by others to increase their own risks.

What comes next for AI oversight

Anthropic plans to release its own internal analysis of the speed of AI progress within the coming weeks. This report will likely provide a different perspective than the METR study, though co-founder Sam McCandlish noted that many of the independent group's conclusions match his own views. He emphasized that the work with METR is separate from the company's ongoing commitment to pacing the frontier.

The broader industry is watching these developments closely as the company prepares for its initial public offering. The success of this IPO could set a precedent for how private AI labs are valued and regulated. As the technology moves faster, the demand for independent, third-party oversight is expected to grow among both regulators and the public.

Moving forward, the focus will likely shift to how companies balance the competitive need for speed with the ethical requirement for safety. If AI continues to drive its own development, the window for establishing meaningful human-centric guardrails may shrink. The upcoming reports from Anthropic will be a key indicator of how the company intends to navigate this tension.

The impact on the AI landscape

For the average user, these developments mean that AI capabilities may begin to jump forward in larger increments rather than steady, incremental updates. As models become better at self-improvement, the complexity of the underlying systems may become harder for non-engineers to interpret. This creates a situation where the