Anthropic Admits AI Now Authors Majority of Its Code at Scale

Anthropic publishes detailed internal software engineering metrics, documenting how artificial intelligence now automates the development of its own successor systems.

The safety-focused intelligence lab published a detailed technical report on June 3, 2026, revealing that automated models now write the vast majority of its core software. According to internal metrics, the Claude model family authored over 80 percent of all code merged into the company’s codebase during May 2026. The exponential scaling in developer productivity enabled Anthropic’s human engineers to ship roughly eight times more code per quarter than they completed in 2024, demonstrating a profound shift in software engineering economics.

The documentation traced the rapid evolution of autonomous agents from simple code assistants to systems capable of managing long-horizon research projects. Claude’s success rate on open-ended software engineering benchmarks rose to 76 percent in May 2026, representing a 50 percentage point increase over a six-month period. On targeted optimization benchmarks, the models went from achieving modest performance gains to delivering 52-fold speedups in specialized training experiments, reducing the timeline required to train new architectures.

The company highlighted the compressed timeline and its broader consequences in an official statement on X. “Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor. It’s happening faster than we thought, and the implications deserve greater attention,” it said.

Anthropic Chief Executive Officer Dario Amodei clarified that full recursive self-improvement remained out of reach despite the automated code volume. He pointed out that independent machines still lacked the high-level judgment and goal-setting capabilities required to train future networks without human supervision. In the published advisory, the research team emphasized that human developers retained total veto power over all core code merges.

The rapid compression of development timelines highlighted profound safety concerns within the research community. While autonomous coding accelerated breakthroughs in scientific research, the prospect of self-improving software raised immediate alignment challenges. Industry experts noted that the ability of a model to optimize its own successor created a feedback loop that could quickly outpace human oversight, creating a strategic race between capabilities and safety verification protocols.

The disclosure prompted other leading artificial intelligence laboratories to evaluate their own automated development practices. Industry analysts noted that if other frontier labs adopted similar levels of automation, the rate of algorithmic progress would decouple entirely from human constraints. The resulting trend suggested that future model iterations would be designed and trained almost exclusively by earlier generations of the same software.

ChainStreet’s Take

Anthropic’s release of internal engineering data proved that the automated feedback loop of recursive self-improvement was already active. By showing that Claude generated 80 percent of its own codebase, the lab demonstrated that artificial intelligence was no longer just a productivity utility, but the primary engineer of its own future. Ultimately, the rapid automation of frontier research proved that the time remaining to solve the alignment problem was compressing far quicker than regulators or developers previously believed.