Anthropic Says Claude Now Leads 26% of AI Research Behind Its Own Successor

Anthropic says Claude now leads 26% of the company’s model research and development, marking a major shift in how the lab builds new AI systems. The model can complete most tasks from a high-level prompt through to an end result, while human researchers retain supervision, the company said. About 90% of Anthropic’s research and development involves some form of collaboration with Claude.

The distinction between leading work and collaborating on it matters. Leading means Claude handles large portions of a project end to end. Collaboration includes narrower contributions, such as writing code, analyzing results, preparing experiments, or helping researchers reason through technical problems. ABC News described the system as still dependent on human oversight, while Bloomberg identified the 26% figure as the clearest measure yet of Claude’s role inside Anthropic.

Claude is helping build its successor, but Anthropic says the process has not reached fully autonomous self-improvement. Researchers set goals, review outputs, and remain responsible for deciding which work enters the model-development pipeline. The company’s description places Claude inside a supervised engineering workflow rather than outside human control.

That boundary is central to the safety question. Anthropic’s blog describes a path toward recursive self-improvement, in which an AI system designs and develops a more capable successor, but says the company has not reached that point and that the outcome is not inevitable. The Washington Post framed the development as Claude taking over work involved in building its own successor, while ABC News emphasized the continuing human role. The gap between those descriptions reflects a system that is increasingly capable, yet still managed by people.

The 26% figure measures more than chatbot adoption. It indicates that Anthropic is using its own models as active research instruments and production tools, compressing parts of the cycle required to design, test, and refine new models. If the figure rises, the company’s development capacity could expand without a matching increase in human labor for every technical task.

Claude’s broader product profile helps explain why it can contribute across the pipeline. Pluralsight describes Claude as a model family, developer platform, coding agent, and productivity tool, while Coursera presents it as a general-purpose generative AI system used by individuals and teams. Those capabilities give Anthropic a model suited to code-heavy research work, though the internal figure applies specifically to the company’s own development process. Bloomberg’s coverage places the claim within an industry debate over how much decision-making AI systems should receive.

Anthropic’s announcement links faster model development with a harder safety problem: the systems helping build future models are themselves difficult to fully understand. The company’s research organization has published data on AI-assisted development while warning that recursive self-improvement could arrive before institutions are prepared for it.

That concern reaches beyond productivity. The New Yorker described Anthropic researchers examining Claude’s internal representations, running psychological experiments, and trying to understand how the system forms responses. Interpretability work becomes more consequential when a model participates in the creation of its successors, because researchers must evaluate both the model’s outputs and the methods it uses to produce them. Claude’s role therefore creates a paired challenge: extracting more useful work while keeping the system’s behavior legible enough to audit.

Anthropic’s internal use of Claude strengthens the company’s position in the competition with OpenAI and other AI developers. Claude already operates as an assistant, coding system, developer platform, and enterprise tool, giving Anthropic a direct feedback loop between product use and model research. Its work on the next generation adds another layer to that relationship.

The development also raises expectations for rival labs. A higher share of AI-assisted research can shorten iteration cycles, improve engineering throughput, and make access to computing resources more valuable. Anthropic’s claims remain company-reported, but they provide a concrete benchmark for measuring internal automation. Bloomberg’s reporting and ABC News’s account align on the core figures, while educational explainers from Pluralsight and Coursera place Claude within the wider market of competing AI assistants.

Anthropic’s next test is whether Claude’s expanding role produces measurable gains in the successor model without weakening oversight. The company will need to show how it defines “leading” work, how researchers review Claude’s contributions, and which safeguards govern experiments that affect future models.

The trend also gives developers and enterprise buyers a practical signal: AI systems are moving from tools that generate outputs to systems that participate in the work of creating other AI systems. Claude has not reached autonomous recursive self-improvement, but Anthropic says the direction is visible. The next milestones will involve greater task ownership, stronger evaluation methods, and clearer evidence that human supervision remains effective as model capability grows.