Anthropic AI research pace doubles as METR report signals automated R&D breakthrough

What the METR report reveals on AI acceleration

Anthropic has approached or crossed its internal threshold for "dramatically accelerated progress", driven by AI systems automating the company's own research and development. An assessment by non-profit evaluation organisation METR found that between October 2025 and October 2026, progress in AI capabilities at Anthropic ran approximately twice as fast as over the preceding twelve months.

The speedup was measured using an internal variant of the Epoch Capabilities Index. According to METR's findings, AI automation accounted for roughly 80 per cent of the observed acceleration by October 2026, with staff reporting that internal AI tools provided up to a 12-fold uplift in research labour efficiency.

The findings coincide with broader industry efforts to measure autonomous AI capabilities in practical domains. OpenAI has published evaluations measuring how AI agents manage routine wet-lab science tasks, assessing their capacity to execute experimental protocols and assist physical research workflows alongside automated software development.

Inside Anthropic's scaling policy threshold

Anthropic's Responsible Scaling Policy defines dramatic acceleration as a sustained doubling of overall capability growth compared to baseline rates before AI contributed significantly to research. Under the policy, this threshold is triggered when speedups are plausibly attributable to AI automating research and engineering rather than hardware expansion.

To sustain double the pace of capabilities growth, researcher efficiency must expand far faster than baseline hiring permits. Researchers at Anthropic reported that internal model assistance effectively allowed individuals to accomplish in one day work that previously required ten days, overcoming diminishing returns in algorithm design.

Safety implications and security risks

The rapid compounding of research progress has raised immediate concerns regarding safety controls and infrastructure security. Under earlier versions of Anthropic's policy, reaching high-tier research automation milestones mandated ASL-4 security standards, designed to prevent model-weight theft by state-level adversaries.

However, security protocols across frontier labs remain under scrutiny following recent security breaches by amateur groups at both Anthropic and OpenAI. Analysts note that without robust model weight protection, universal logging, and verified alignment techniques, rapid recursive acceleration could outstrip existing governance frameworks before risks are adequately mitigated.

What happens next will depend on whether hardware compute constraints or algorithmic bottlenecks slow down the feedback loop, or whether frontier labs can implement security controls capable of matching accelerated development timelines.