Anthropic: Doubling the Pace of Internal AI Advancements

Researchers affiliated with METR describe an acceleration of approximately two times the capabilities of Anthropic between October 2025 and October 2026, measured by an internal index. Anthropic, for its part, indicated in its S-1 that it could soon reach the acceleration threshold outlined in its scaling policy, defined as a pace twice as fast as the previous most sustained rate. A large portion of the evidence put forward by the authors has been redacted.

Shortened timelines and a major signal on risk

The report is presented as the most significant update on AI-related risk in the past two years, more so than the Hugging Face attack. According to this reading, the observed dynamics suggest shorter horizons than those held by many external researchers, including the AI 2027 team. The subject is presented as much more important than it appears at first glance. It is also suggested that if AI doubles the overall pace of capabilities, the acceleration of human researchers should be even more pronounced.

A leap measured by the AECI and largely redacted evidence

According to the authors of the report, between October 2025 and October 2026, advancements in capabilities at Anthropic have progressed at a rate approximately twice that observed in the previous twelve months, an assessment made using an internal version of the Epoch Capabilities Index, the AECI. Many supporting pieces of evidence have been removed from the published version, and these withdrawn documents would represent the largest share of information considered most enlightening by the authors. Colleagues from METR have conducted a thorough investigation in recent months to assess the extent to which Anthropic is accelerated by its AI systems. The simultaneous review of this report and the company's policy suggests that a major acceleration threshold has already been surpassed. The authors believe that AI-driven automation is likely the main cause of the current acceleration in capability increases, while noting that their detailed explanations have been omitted.

Anthropic's acceleration threshold and mention in the S-1

The latest version of Anthropic's scaling policy, in effect since July 2026, sets the crossing of the threshold when the overall capability progress reaches a pace twice as fast as the previously fastest sustained rate before a significant contribution of AI to R&D. Anthropic indicated earlier this week in its S-1 that it could imminently reach this stage. It is also suggested that the automation of AI R&D by AI itself likely plays a more decisive role than merely increasing computing power or other inputs.