Via tomshardware.com
The AI company says its own model writes the vast majority of code merged into its codebases, with engineers shipping eight times more code per day than historical averages.
Anthropic just published an internal report showing that Claude, its flagship AI model, authored over 80% of code merged into the company’s production codebases as of May 2026. That figure was in the low single digits before the February 2025 research preview of Claude Code, the company’s agentic coding tool.
The numbers behind the shift #
In Q2 2026, Anthropic engineers merged approximately eight times more code per day compared to the company’s averages from 2021 through 2024. Claude Code hit research preview in February 2025 and reached general availability by May 2025.
Perhaps the most striking metric in the report: Claude achieved a 76% success rate on complex engineering tasks by May 2026. That represents a 50-point jump over just six months.
The report also highlights an automated Claude code review system that checks all proposed changes to Anthropic’s codebases. That reviewer has caught approximately one-third of bugs from prior incidents involving claude.ai.
Recursive self-improvement, minus the sci-fi #
Anthropic’s report uses the phrase “recursive self-improvement” to describe what’s happening internally. In practice, this means Claude independently manages technical implementations and experiments while human engineers focus on strategic decisions and code reviews.
Anthropic’s report notes that code quality from Claude has reached parity with human-authored code.
What this means for software engineering #
Anthropic was founded in January 2021 by Dario Amodei and Daniela Amodei, along with other former OpenAI researchers.
The report also surfaces a tension that every company adopting AI coding tools will eventually face. Anthropic acknowledges the risks around control and alignment when AI systems take on this much of the engineering workload. The automated Claude reviewer catching one-third of bugs from past incidents is encouraging, but it also means two-thirds of those bug patterns weren’t caught by the automated system.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our