An 11-week audit by AI safety organization METR reveals that Anthropic is significantly accelerating its internal research pace by deploying its own unreleased AI models. The report highlights that Anthropic’s models are advancing roughly twice as fast as they did a year ago, compressing two years of historical research progress into the past twelve months.

METR confirmed that the acceleration stems primarily from AI capabilities rather than simple compute scale expansion. Internal survey data indicated that researchers utilizing advanced internal systems reported being approximately 12 times more productive than without AI assistance—doubling self-reported efficiency gains logged just months prior. Over the past six months, Anthropic completed training on three unreleased internal systems codenamed Model 2, Model 3, and Model 4.

While most current gains rely on human-AI collaboration, METR warned that fully autonomous R&D contributions may soon play a larger role. The report noted an absence of clear diminishing returns, raising questions about fast feedback loops in frontier AI labs as Anthropic prepares for its public offering.

Why it matters

  • Recursive AI self-improvement is shifting from theoretical research to practical reality inside top tier frontier laboratories.

  • AI-assisted developer productivity for AI researchers is scaling exponentially, giving leading labs a compounding advantage over smaller competitors.

  • Safety auditors face visibility challenges when frontier labs redact key performance data prior to public disclosure.

Source: theverge.com