Anthropic says Claude now leads a quarter of work building its next AI models

Anthropic disclosed on Thursday that Claude now leads 26% of the artificial intelligence research and development work within the company, as part of new measures to regularly publish insights for outsiders. The startup, based in San Francisco, reported that AI collaborated with humans on over 90% of the research work at the company as of August, according to a blog post.

Anthropic clarified that Claude is not fully autonomous in any aspect of the measured work, but its contribution to AI development has surged rapidly, from 1% in March, according to a scale established by Epoch AI, an independent nonprofit tracking the technology. The company's figures provide a detailed public view of how leading firms are advancing AI amid concerns about its potential to autonomously develop behaviors diverging from human intentions.

On Wednesday, OpenAI, Anthropic's rival, announced plans to regularly publish reports on unexpected or unauthorized AI behavior, disclosing six instances of such behavior. Anthropic revealed that around 30,000 AI agents were engaged in research and engineering tasks on its primary internal platform in August, with every action screened before execution.

Over a billion decisions were made in that month, with roughly one in 47,000 blocked. The firm also shared that approximately 6% of the computing power used for AI research went to safety work in a July sample week, rising to 12% for research conducted by AI. The company emphasized that these figures are conservative as computing power utilized equally for safety and capability was counted as capability work.