Claude Code Team Reveals Internal Practices: About 90% of Code Written by Claude, Anthropic Merges Over 30,000 Changes Internally in a Day
Anthropic has officially opened the Code Review feature of Claude Code from research preview to Claude Team and Enterprise plan users.
The review speed has increased by about three times, with most reviews reduced from about 20 minutes during the preview period to 6 to 10 minutes, while capturing more critical defects. Results are presented in a concise manner detailing issues, triggering conditions, and solutions. Administrators can choose Quick, Standard, or Deep review levels for each repository to match different codebase needs and costs.
Eligible Enterprise organizations can receive at least $10,000 in Code Review credit before January 15, 2027, valid for 30 days after receipt. This feature automatically runs on pull requests through a GitHub application and leaves comments on code lines.
The Claude Code team has revealed its internal practices: about 90% of the code is written by Claude. From June to September, the number of merged pull requests in the Claude Code repository more than doubled, while the retraction rate decreased by over 40%.
The team's goal is to merge most pull requests with at most one person involved, without increasing the retraction or incident rate. The review system uses a multi-agent division of labor, with one group looking for defects and another group assessing the authenticity of defects, while human approval is set based on risk levels.
Mechanically, enterprise clients purchase this service due to review bottlenecks, and Anthropic promotes adoption through credit limits and tiered pricing. As code generation speeds up, review becomes the main constraint, allowing teams using Code Review to reduce manual review time, putting pressure on traditional purely manual review processes.
Supplementary data shows that Anthropic's internal daily merged changes increased from about 2,500 in May to over 30,000 in September, with review costs billed based on token usage, averaging about $15 to $25 per review.
Source: Public Information
ABAB AI Insight
Anthropic launched the Code Review research preview in March 2026 and has been running a multi-agent system on nearly all internal pull requests. It was previously disclosed that Claude wrote over 80% of merged production code, with engineers' daily merge volume increasing about 8 times compared to 2024. Recently, it was further revealed that about 90% of the Claude Code team's code is generated by the model, achieving a doubling of merge volume from June to September while the retraction rate decreased by over 40%.
In terms of capital pathways, Anthropic is productizing its high-output internal review process, charging through Team and Enterprise plans, and providing eligible enterprises with $10,000 to $50,000 in credit to accelerate adoption. Resources are shifting from model training and code generation capabilities to review agent deployment and GitHub integration, motivated by transforming the review bottleneck into a paid service while validating product reliability with its own data.
Similar cases include GitHub Copilot expanding from code completion to review and automation processes, as well as the transition of early static analysis tools to AI-assisted review. Currently, it is in a phase of expanding from code generation dominance to a closed loop of generation and review, with institutional clients beginning to incorporate multi-agent review into standard development pipelines.
Essentially, this is a technological substitution, where the mechanism is that once the model's code output speed exceeds manual review capacity, a proxy system must handle the review volume; otherwise, the output bottleneck shifts to the manual stage, leading to an inability to synchronize improvements in merge speed and quality. The industry chain is shifting from a labor-intensive manual review model to a proxy-assisted and tiered human intervention model.
ABAB News · Cognitive Law
When generation speed exceeds review speed, the bottleneck shifts from writing code to reviewing code.
Once agents write code, agents must review code to maintain output.
A doubling of merge volume while the retraction rate decreases is evidence of an effective review system.