Three unnamed internal Anthropic models were recently leaked, having completed training in July, August–September, and September–October of this year, respectively. Among them, the strongest model—Model 4—scored approximately 3.9 points higher than the previous Opus 5.5 on the AECI evaluation, roughly equivalent to three months of Anthropic's capability progress. Since October 2025, the AECI growth rate of Anthropic's models has nearly doubled, with AI-assisted R&D cited as the primary driver of this acceleration. AI assistance has already cut the time from idea generation to running experiments by about 80%, and code is no longer the main bottleneck. Co-founder Sam McCandlish confirmed that the report was stripped of a large amount of commercially sensitive information and previewed that an analysis of the company's own R&D progress will be released in the coming weeks. Although Anthropic's leadership has publicly called for slowing the pace of AI capability improvement, internal model iteration has not slowed; the core motivation behind its push for technological leadership is the hope that AI technology will be held by entities that prioritize safety.