Claude Leads Development of Next-Gen AI Models at Anthropic

Anthropic said on Thursday 17 September that Claude is helping to develop the next, more capable version of Claude, marking a sharp increase in the AI system’s role inside the company’s research operation.

According to Anthropic, Claude was leading 26% of model research and development work by August 2026. That means it could complete most of a task “end-to-end from a high-level prompt” while under human supervision, according to the Associated Press. It was also involved in roughly 90% of R&D work in a more collaborative capacity, completing large portions of tasks under close human direction.

The company said the share of work Claude led was zero in February. It also said about 30,000 AI agents were carrying out research and engineering work internally as of August.

Not autonomous self-improvement

Anthropic is drawing a firm distinction between today’s workflow and fully autonomous AI systems that can design their own successors. It defines that possible future as recursive self-improvement, but says: “We are not there yet, and recursive self-improvement is not inevitable.”

Humans still set the broader direction, Anthropic said. “The comparative advantage of humans as of right now is still in seeing the bigger picture and thinking beyond the confines of the immediate task.”

Claude’s growing role in coding

Anthropic says more than 80% of code merged into its codebase had been authored by Claude by May 2026, using a measure limited to code merged into production. It also said the typical engineer was merging eight times as much code per day in the second quarter of 2026 as in 2024, while cautioning that code volume is not a direct measure of quality.

The company uses an automated Claude reviewer to inspect proposed changes for bugs and security flaws. Its retrospective analysis found that applying this review to every code change could have caught about one-third of the bugs behind previous incidents on claude.ai before production.

Anthropic says Claude-written code was below human-written code at the company in late 2025, but is now roughly at parity. That remains the company’s own assessment rather than an independent comparison.

Why it matters

AI assistants are already used by consumers to write, analyse and code. Anthropic’s disclosure suggests the same systems are becoming part of the machinery used to build subsequent generations of AI, potentially shortening the gap between model releases and improvements.

For users, that could eventually mean more capable tools arriving faster. But the headline figure should not be mistaken for an AI independently rebuilding itself: Anthropic says Claude remains supervised, while people continue to choose the problems, set the goals and assess the wider picture.