On September 18, Anthropic announced that Claude is helping build the next, more capable version of itself. According to the company, Claude currently leads 26% of Anthropic's model research and development -- meaning it can complete most of a given task end-to-end from a high-level prompt, while still operating under human supervision rather than fully autonomously. More broadly, Anthropic says roughly 90% of its R&D now happens "in collaboration" with Claude, meaning the model handles large chunks of work under close human direction rather than starting from scratch itself.
On the product side, Claude Fable 5.1 released September 1, 2026, positioned for demanding reasoning, long-running agents, coding, multistep research, and document-heavy professional work. It ships a 1-million-token context window, up to 128K output tokens, and what Anthropic describes as adaptive thinking that stays engaged throughout a task rather than switching on only for harder sub-problems.
Worth sitting with both facts side by side: a lab reporting that its own model does a growing, quantified share of that lab's own model R&D, disclosed the same month its CEO published an essay explicitly asking the industry to slow the pace of capability gains. Whether Anthropic keeps publishing that specific self-improvement percentage next quarter, or lets it quietly stop being a headline number, is worth checking back on.
Anthropic quantifying Claude's contribution to its own R&D (26% leading, ~90% collaborative) is a genuinely new kind of disclosure -- the number to watch isn't this one, it's whether the company keeps reporting it as that percentage climbs.