Relying on Claude, more than 80% of Anthropic's code is now generated by AI, engineering productivity has multiplied several times, and AI autonomous iteration and recursive self-improvement are progressing faster than expected. In its latest technical blog, the U.S. large-model company Anthropic disclosed measured data in When AI Builds Itself, forecast three industry trajectories, and made a rare appeal: reasonably slow the pace of frontier AI R&D to give society time to improve supporting safeguards, which may benefit the safe development of all humanity.
Anthropic's latest technical blog, When AI Builds Itself
Anthropic said that as of this May, more than 80% of Anthropic's code was written by Claude. Before the release of Claude Code, that figure was still in the single digits. At the same time, the average amount of code Anthropic engineers delivered per quarter was eight times the level seen from 2021 to 2025.
“During the first four years after Anthropic was founded, from 2021 to 2024, the number of code lines merged per engineer per day remained basically stable; in 2025, as Claude began running code rather than merely suggesting that engineers copy and paste it, that number started to rise.”
Anthropic said that in 2026, as models began operating autonomously over longer time horizons, the growth rate accelerated again. Data show that in Q2 2026, the amount of code merged per day by an average Anthropic engineer was eight times that of 2024.
With code written by Claude being “good” and continuously improving, what will future work look like? Anthropic says there may be at least three different scenarios:
First, although exponential growth may stall, AI capabilities are now widely diffused. Anthropic says we are still in the early stage of rolling existing models out across broader parts of the economy, but the constraints on AI progress may not be the model itself; they may lie in supply chains, chip manufacturing speed, grid expansion speed or interconnect bandwidth.
Second, the efficiency of AI labs will keep improving, and AI development will become highly automated, while humans remain responsible for setting research directions and evaluating results. Over time, organizations that use AI systems will become more efficient, and everyone's productivity inside those organizations will rise significantly. In companies like Anthropic, the human role will also change. People will work with AI systems to scale research and jointly build the systems needed to verify whether AI can be trusted.
Third, AI systems themselves will gain full recursive self-improvement capabilities and begin building their successors. If technological progress continues and AI systems can develop creativity comparable to humans, then AI systems are likely to design and improve themselves.
Anthropic believes that in such a world, the pace of AI development will depend entirely on the compute resources required by AI systems (or the speed at which various efficiency gains are discovered in training or inference), while the role of humans in AI system development will diminish substantially. “Most of our energy may shift to supervising, verifying and validating the ‘virtual laboratories’ in which AI systems operate.”
However, Anthropic said it is still too early to judge which outcome is most likely, but AI already helping build AI is a reality.
What should we do? In this technical blog, Anthropic says that if the pace of development can be effectively slowed, giving us humans more time to respond to its enormous impact, that may be a good thing. But if slowing down instead allows the least cautious players to catch up technically, it could ultimately threaten everyone's safety.
“We believe it would be very beneficial for the world if it could choose to slow down or temporarily pause frontier AI R&D so that social structures and coordination research can keep pace with technological progress,” Anthropic said.
The article was co-authored by Marina Favaro and Jack Clark, with Clark being a co-founder of Anthropic and Favaro working on matters related to the Anthropic research institute.
Notably, Anthropic's public warning about AI risk came amid a rapid rise in the company's valuation and preparations for going public. On June 1, Anthropic announced on its official website that it had secretly submitted a draft Form S-1 registration statement to the U.S. Securities and Exchange Commission (SEC) for a common stock IPO. The company also emphasized that the number of shares and pricing had not yet been determined, and that the IPO would depend on market conditions and other factors. According to the Financial Times, Anthropic has also filed IPO-related documents and is seeking to list at a valuation above US$1 trillion, making it, together with OpenAI and SpaceX, one of the focal points of the new wave of tech giant listings. If successful, Anthropic could become one of the most closely watched public listings among the next generation of AI companies, and would also provide a reference for how capital markets price generative AI companies.
