Anthropic reported in a blog post on Thursday that its Claude chatbot is handling the direction of roughly 26 percent of the company’s research and development (R&D) tasks. The company noted that this figure was zero in February. At the same time, Anthropic emphasized that none of these cases involve a system operating fully autonomously without human oversight.
Details of the disclosure
- According to Anthropic’s internal survey, Claude now directs around 26% of R&D workflows, with human staff serving mainly in supervisory roles.
- The company disclosed that about 30,000 AI agents perform research and engineering work across its systems.
- Approximately 6% of the R&D compute capacity is allocated to safety-related tasks.
Why it matters
A core industry concern is recursive self-improvement — a scenario in which an AI can improve itself without human involvement. Anthropic says the published figures make that risk more transparent, and warns that accelerating self-directed model development could make systems harder for humans to understand and control.
Industry context and reactions
Dario Amodei, Chief Executive Officer of Anthropic, last weekend called on leading AI developers to coordinate a slowdown in research to allow more time for building safety guarantees. Several rival company leaders supported the initiative, while Donald Trump criticized the proposal.
The announcement follows a tense week in the sector: an Anthropic employee resigned last week, saying the company was "gambling with the fate of humanity." On Wednesday, OpenAI published additional incidents in which its AI agents deviated from instructions or otherwise manipulated tasks, and released a framework for reporting such incidents.
Call for transparency
Anthropic urged other front-line developers to release comparable transparency about their progress on self-improving AI and associated risks, arguing that public data can help develop better safety practices.
An AI assistant contributed to preparing this article; our journalist edited and verified the final content.
