Anthropic is calling for a more cautious approach to artificial intelligence development as its Claude models become increasingly capable of performing software engineering tasks.
According to the company, Claude now generates more than 80% of Anthropic’s production code, a dramatic increase from only a few percentage points before the launch of Claude Code in early 2025.
The company also reported that a typical Anthropic engineer merged roughly eight times more code per day during the second quarter of 2026 compared with 2024.
Claude Approaches Human-Level Coding
Anthropic stated that by the end of 2025, code generated by Claude was still somewhat inferior to human-written code.
Today, the company believes Claude’s coding performance is roughly on par with human engineers in many situations and expects the model to become significantly better within the next year.
Executives emphasized that the volume of generated code should not be confused with productivity, noting that software quality and real-world outcomes remain the most important metrics.
The rapid improvement of coding models is increasingly transforming software development workflows across the technology industry.
Concerns Over Self-Improving AI
One scenario highlighted by Anthropic involves AI systems that begin automatically improving and rewriting themselves with limited human intervention.
The company acknowledged that the long-term consequences of recursive self-improvement remain uncertain and argued that society should spend more time studying the potential risks before accelerating deployment further.
Anthropic plans to engage policymakers, regulators, researchers, and other AI laboratories in discussions about governance and safety measures.
The debate reflects growing divisions within the AI industry, where some companies advocate for faster deployment while others emphasize caution around increasingly autonomous systems.
The broader takeaway is that leading AI developers are beginning to confront questions not only about how quickly artificial intelligence can improve, but also whether safeguards can keep pace with the technology’s accelerating capabilities.
