The latest

Anthropic is increasingly using Claude to help develop future AI models, shifting its role from a researcher’s tool toward a system that executes complex parts of the development process. By August, Claude could complete most of about 26% of model research and development tasks from high-level instructions, while people continued to supervise. Anthropic said Claude was not fully autonomous, but its expanding contribution allows researchers to oversee substantially larger volumes of work.

Details

• Rapid expansion: The 26% share was reached in August, around six months after Claude began leading a smaller portion of Anthropic’s research. In the company’s definition, leading a task requires the model to carry out most of the work after receiving an overarching instruction, with human oversight continuing throughout.
• Embedded collaboration: Around 90% of Anthropic’s model R&D activities now involve work with Claude. The system can handle substantial volumes across those activities, though the level of involvement differs from the 26% of tasks it leads and remains closely supervised by staff.
• Agent scale: Anthropic said approximately 30,000 AI agents were carrying out research and engineering work in August. Their assignments span the types of multi-step activity frontier models increasingly perform, moving beyond assistance with individual coding, data-analysis or idea-generation tasks toward more complete workflows.
• Economic leverage: Greater automation of software engineering, experimentation and testing could shorten development cycles and enable relatively small research teams to manage more work. That possibility adds a new competitive factor: how effectively a model improves the process used to build its successor.
• Recursive prospect: More capable systems helping create still more capable models could generate a development feedback loop associated with recursive self-improvement. Anthropic said its measurements do not indicate precisely how close Claude is to continuous or recursive improvement, keeping the current milestone distinct from fully autonomous model development.
• Transparency campaign: The company is urging other AI developers to publish comparable metrics through a common methodology. It argues that tracking progress over time and across laboratories would narrow the information gap between frontier companies and governments, independent researchers and the public as spending rises on chips, computing infrastructure and data centers.

Between the lines

The competitive contest is broadening beyond which laboratory can train the most capable model. A system that performs more of the research and engineering behind its successor could give its developer an advantage in speed, scale and use of human expertise.

What’s next

The next measurable indicator is whether Anthropic’s future disclosures show Claude leading more than the August level of 26%, and whether other frontier laboratories adopt the proposed common methodology and release directly comparable figures over time.