Archive · 18 Sep 2026
Anthropic Says Claude Leads 26% of Next AI Development Work
Anthropic says its Claude AI system is helping build newer AI systems.
In August, Claude led 26% of the company’s research and development work.
That share was only 1% in March.
AI also worked with people on more than 90% of the company’s research.
About 30,000 AI agents were doing research and engineering tasks.
Anthropic says the agents’ actions were screened before they happened.
Only about one decision in 47,000 was blocked during a month with more than one billion decisions.
Researchers are concerned that more independent AI could become harder to monitor and control.
Claude led 26% of Anthropic’s internal artificial intelligence research and development work in August, up from 1% in March.
AI collaborated with humans on more than 90% of Anthropic’s research work as of August.
About 30,000 AI agents conducted research and engineering work on Anthropic’s main internal platform in August.
Anthropic said agents’ actions were screened, with about one in 47,000 decisions blocked among more than one billion decisions.
About 6% of Anthropic’s AI research computing power went to safety work, rising to 12% for research conducted by AI itself.
- Who
- Anthropic, its Claude AI system, human researchers, and AI agents.
- What
- Anthropic reported that Claude led 26% of its internal AI research and development work, while AI collaborated with humans on more than 90% of research.
- When
- The figures cover August, with safety-computing data from a sample week in July; Anthropic announced the measures on September 17.
- Where
- The work took place inside Anthropic, a San Francisco-based company, including on its main internal platform.
- Why
- Anthropic said it will publish the measures regularly to show outsiders how quickly AI is contributing to the development of newer AI systems.
Anthropic’s Transparency and Oversight Claims
Researchers’ Safety Concerns
AI involvement in development
Anthropic’s Transparency and Oversight Claims
Anthropic says regularly publishing measurements will give outsiders a clearer view of how AI is helping develop the next generation of AI systems.
Researchers’ Safety Concerns
Researchers warn that increasingly autonomous AI agents may develop behaviors that diverge from their creators’ intentions and become harder to monitor or control.
Agent oversight
Anthropic’s Transparency and Oversight Claims
Anthropic says every action by its research agents was screened before execution and that Claude was not fully autonomous in any measured work.
Researchers’ Safety Concerns
The use of about 30,000 agents and more than one billion decisions in August highlights concerns about monitoring and controlling AI systems at scale.
Safety investment
Anthropic’s Transparency and Oversight Claims
Anthropic reported that 12% of computing power for AI-conducted research went to safety work and said its figures were conservative because computing serving both safety and capability was counted as capability work.
Researchers’ Safety Concerns
Safety work represented 6% of overall AI research computing in the sample week, while researchers continue to warn that more autonomous AI may become harder to control.
- Claude’s research contribution
- 26% of Anthropic’s AI research and development work in August
- Increase since March
- Claude’s contribution rose from 1% in March
- Human-AI collaboration
- AI collaborated with humans on more than 90% of Anthropic’s research work as of August
- AI agents in use
- About 30,000 agents conducted research and engineering work on Anthropic’s main internal platform in August
- Blocked decisions
- About one in 47,000 decisions was blocked out of more than one billion decisions in August
- Safety computing
- About 6% of computing power used for AI research went to safety work in a July sample week, compared with 12% for research conducted by AI
- Autonomy status
- Anthropic said Claude was not operating fully autonomously in any measured area
