Anthropic says Claude now performs 26% of the work behind its next models
The figure offers a rare view of AI helping build AI—and raises new questions about productivity, oversight and feedback loops inside frontier labs.
The story
Claude is now responsible for roughly 26% of the research and development work used to build Anthropic’s next AI models, the company said in an account reported by Reuters.
AI-assisted coding, experiment design and analysis can compress development cycles. It can also create correlated errors if teams depend on the same models to produce and review important work.
The number is a sign of real workflow change, but not a complete productivity measure. The important evidence will be whether human researchers catch failures, improve scientific quality and document where automated contributions influence safety-critical decisions.
INNOVOX analysis
Using AI to build AI can compound productivity, but it can also amplify shared blind spots. Labs need strong separation between generation and evaluation so the same model family does not silently validate its own errors.
What to watch
More useful reporting would separate coding, analysis and experiment design, and show how automated contributions are reviewed in safety-critical work.