Anthropic Discloses Claude Now Leads 26% of Its Own Model Development
The AMW Read
Anthropic quantifies AI-driven self-improvement inside a top-tier lab for the first time and calls for shared reporting standards, updating its case-study profile and feeding the safety-governance debate without resolving it.
Anthropic Discloses Claude Now Leads 26% of Its Own Model Development
Anthropic said last week that Claude is playing an increasingly central role in building the company's next generation of models. As of August, Claude was the leading contributor on 26% of Anthropic's internal R&D work β up from essentially zero in February β completing most end-to-end tasks from high-level prompts under human supervision, though not yet operating fully autonomously. Anthropic said close to 90% of its R&D work now involves some form of Claude collaboration, with humans directing large blocks of the effort, and that roughly 30,000 agent instances were engaged in R&D engineering as of August. The company plans to bring in outside third-party evaluators to review its safety practices, though it did not disclose how far it believes it is from full recursive self-improvement.
The disclosure follows public warnings from CEO Dario Amodei and other Anthropic leaders about the pace of AI development, and lands the same week Anthropic detailed the embedded-evaluator framework behind its non-exclusive Accenture safety pact and its talks with nonprofit evaluator METR. Anthropic argues that as models increasingly build their successors, the information gap between frontier labs and the public widens, and it wants competitors to adopt a shared, comparable methodology for reporting how much of their own model development is AI-driven β an attempt to set an industry norm rather than let self-improvement metrics stay proprietary.
For enterprises and investors, this is one of the first quantified, first-party benchmarks for how much frontier R&D is already automated at a top lab, and it puts pressure on OpenAI, Google DeepMind, and xAI to either match the disclosure or explain why they won't. Builders using Claude for coding and agentic workflows get a concrete signal of how aggressively Anthropic is scaling internal agentic tooling ahead of its expected IPO.
