Anthropic AI Models Accelerate AI Development: Research Finds Nearly Double the Pace of Progress

Anthropic is increasingly using its own artificial intelligence models to accelerate the development of more advanced AI systems, achieving roughly two years of historical progress in just one year, according to a report by independent AI safety research organisation METR. The findings are based on an 11-week investigation into the company’s internal research and development activities. METR estimates that Anthropic’s model benchmark scores have improved at 1.9 times their earlier rate since October 2025. The acceleration is largely attributed to AI-assisted research rather than increased computing power. The report also found that employees using Anthropic’s most capable internal models reported being approximately 12 times more productive than they would have been without AI assistance. These productivity gains were linked to unreleased models internally referred to as Model 2, Model 3 and Model 4, developed over the past six months.

The findings have renewed concerns about the possibility of AI systems eventually designing and developing their own successors with little or no human involvement, a concept known as recursive self-improvement. METR said it had not identified significant bottlenecks that would necessarily prevent further progress over the next 12 months, although it acknowledged uncertainty in its assessment. The report was based on confidential research, performance data and employee interviews provided by Anthropic. However, the company removed substantial portions of the findings before publication, prompting criticism from METR that the redactions made independent verification difficult. Anthropic co-founder Sam McCandlish defended the decision, saying that revealing the details behind the acceleration could expose commercially sensitive information and provide competitors with a roadmap. The findings have also sparked debate over whether Anthropic’s rapid AI development aligns with its public emphasis on AI safety.