Anthropic Says Claude Now Helps Build Its Successor

Anthropic says Claude is no longer only a product being improved by researchers: it is now helping design and develop the company’s next generation of models. The disclosure, published Thursday, places one of the industry’s most important theoretical risks—recursive self-improvement—inside an operating AI lab, even if the company stresses that the process remains human-supervised.

According to Anthropic, Claude currently “leads” 26% of the company’s AI research and development work. In this context, leading means the model can complete most of a task end-to-end from a high-level prompt, while researchers still set objectives, review outputs, and decide which work is accepted. Roughly 90% of Anthropic’s R&D is described as collaboration with Claude, meaning the model performs substantial portions of projects under closer human direction. (anthropic.com)

What changed

The significance is not that Claude has independently trained and released a successor. Anthropic explicitly says it has not reached that point. Instead, the company is reporting a transition in where human researchers spend their time: less on writing every component directly and more on specifying problems, supervising agents, checking results, and choosing which experiments deserve expensive compute.

Anthropic’s internal figures suggest the shift is already affecting productivity. More than 80% of code merged into its codebase was authored by Claude as of May 2026, while typical engineers were merging about eight times as much code per day in the second quarter as in 2024. Those figures measure output, not necessarily equivalent scientific progress, but they show why AI labs see model-assisted research as a competitive advantage. (anthropic.com)

The company’s broader argument is that AI systems are becoming capable of longer, more open-ended work. Anthropic says the length of tasks models can complete reliably has been doubling roughly every four months, with newer systems moving from short coding fixes toward investigations and multi-step research assignments. That trajectory could make AI a force multiplier for the teams building frontier models themselves. (anthropic.com)

Why it matters

If the trend continues, model improvement could become partially self-reinforcing: better systems help researchers produce the data, code, evaluations, and experiments needed to build still-better systems. That could compress development cycles and widen the gap between labs with access to capable agents and those without them.

It also changes the safety problem. Human review remains present, but reviewers may increasingly evaluate work generated at a scale they cannot reproduce manually. The key question becomes whether oversight can keep pace with the model’s contribution, especially when the model is helping design training methods, evaluations, or safeguards for its own successors.

What remains uncertain

Anthropic’s percentages are company-defined measures, not an independent audit, and “collaboration” covers a wide range of tasks. The announcement does not establish that Claude can autonomously choose research goals, run a full training pipeline, or reliably improve its own underlying capabilities. Those distinctions matter: assisted engineering is already widespread, while fully recursive self-improvement would require much greater autonomy, access, and reliability.

Still, the report marks a consequential threshold. AI labs are beginning to use frontier models not merely to ship software, but to accelerate the research process that determines what frontier models become next.