Anthropic: Claude writes 80% of merged code, engineers bottleneck

Anthropic reported in a paper titled “When AI Builds Itself,” published Thursday, that its Claude model now authors more than 80% of the code merged into the company’s codebase and that engineers are shipping roughly eight times more merged code than in 2024.

The company reported that lines of code merged per engineer per day remained steady from 2021 through 2024, then began to rise in 2025 after a research preview of Claude Code launched in February. Before that preview, merged contributions from Claude were described as being in the low single digits.

Anthropic described a change in how engineers use Claude: the model moved from suggesting snippets for humans to copy and paste to running code, proposing and executing experiments, and applying findings directly to the codebase. The company noted that higher merged output reflects these new workflows but cautioned that lines of code are an imperfect measure of research progress.

The report lays out possible paths for development as AI takes on more engineering tasks. One scenario keeps humans in oversight and validation roles while AI automates routine work. Another scenario envisions systems that design and build more capable successors, a process called recursive self-improvement. The paper states, “Taken far enough, and given enough compute, that trend points to an AI system capable of fully autonomously designing and developing its own successor.” On the company’s X account the authors added, “None of this guarantees recursive self-improvement is on the horizon. It’s not yet clear that Claude is capable of research judgment-of choosing the right problems to work on.”

Anthropic reported recent technical updates, including an upgrade of its flagship model to Opus 4.8 last month and development of Claude Mythos, which the company reported can identify software vulnerabilities and conduct complex cybersecurity research. The report describes a likely shift in human roles toward oversight, validation and verification of an increasingly automated virtual lab run by AI systems, and it suggests that automated research skills could transfer to other scientific fields.

The paper was published amid other industry model releases, including GPT-5.5 and GPT-Rosalind in April, and as Anthropic prepares for a potential public listing. The company emphasized uncertainty about timelines and outcomes and urged further study while documenting the reported changes in how its engineering teams work with Claude.