New report reveals Anthropic's AI research speed doubled with machine assistance

Speeding Up Progress

Anthropic is using artificial intelligence models to build more powerful AI systems, completing two years of historical progress in just one single year. An independent AI safety research organisation named METR published this finding on Thursday, November 5, following an 11-week investigation into Anthropic's internal research and development work. METR found that Anthropic's models are improving roughly twice as quickly as they were a year prior. Researchers confirmed this acceleration is not driven by an increase in training compute. Instead, the majority of the progress comes from AI models acting as research assistants to human scientists, though human staff remain in control of the process for now. METR noted that employees using the company's most capable internal model self-reported a 12-fold increase in work productivity compared to working without AI help. This figure doubles the gains staff reported a few months earlier while using the Claude Mythos 5 model.

Unreleased Models and Redactions

The acceleration is largely powered by unreleased internal AI systems built over the past six months. These models carry internal designations such as Model 2, Model 3, and Model 4, with training completed across recent months. METR also identified an older unreleased system from 2022 called Unreleased 2022. The IPO-bound startup granted METR access to confidential research data and employee interviews for this assessment. However, Anthropic exercised its agreement right to redact large portions of the final text before public release. METR stated that the startup redacted a majority of the most informative evidence, leaving the public to rely heavily on the organisation's own judgment. Responding to this, Anthropic co-founder Sam McCandlish explained that the company redacted details to protect commercial sensitivities and avoid handing out a development roadmap to rivals. Meanwhile, rival company OpenAI has separately stated it plans to build an AI researcher model by March 2028.

Context

This report brings attention to recursive self-improvement, a theoretical stage where artificial intelligence can autonomously design and upgrade its own successors without human input. While technology companies race to build more advanced models, safety researchers continually debate whether governance frameworks can keep pace with accelerating development timelines. Understanding these internal lab dynamics helps policymakers and the public gauge how quickly frontier technology capabilities are actually advancing behind closed doors.