Anthropic CEO Discusses Share of AI-Generated Code

Anthropic CEO Dario Amodei said in March 2025 that AI would be writing 90% of code within three to six months, in remarks Tech Trenches and Daring Fireball both later revisited. Anthropic engineer Boris Cherny posted on X in late December 2025 that in the last thirty days, 100% of his contributions to Claude Code were written by Claude Code, and Tech Trenches reports an Anthropic product leader separately describing some teams as near-100% AI-written. Tech Trenches' own contribution was a review of a leaked portion of the Claude Code codebase, which it says shows a single 3,167-line function, regex-based sentiment heuristics, and a documented bug that wasted 250,000 API calls per day, evidence it uses to question whether heavy AI authorship correlates with clean code.

Redwood Research's analysis, published in October 2025, argues the definitional basis for these percentages is the real story: metrics like lines merged, commits authored, characters typed, or any code judged "at all useful" produce very different headline numbers from the same underlying activity. Redwood estimates Anthropic's company-wide merged-line fraction is likely closer to 50%, well below the 90% figure, while acknowledging individual teams or engineers can plausibly reach much higher fractions depending on how contributions are counted.

Writing about the same underlying claim in March 2026, Daring Fireball's John Gruber offered a different reframe: the real shift may not be that AI replaced human-written lines one-for-one, but that AI code generation is producing services, apps, and libraries that would not have been written at all otherwise, meaning the human-written line count may be roughly flat even as AI-generated volume grows far faster. Gruber cited GitHub data in a May 2026 update showing commit volume up roughly 14x year-over-year. Neither framing resolves the underlying measurement problem Redwood Research raised, but both point to the same practical takeaway: aggregate percentages without a stated denominator understate how contested this question still is.

Anyone evaluating AI-code claims, whether a vendor's marketing figure or an internal engineering metric, should ask what specifically is being counted (lines merged versus commits versus any AI-touched code), whether the figure is company-wide or cherry-picked from a high-adoption team, and whether code volume and code quality are being conflated. Tech Trenches' leaked-codebase findings are a reminder that a high AI-authorship percentage says nothing on its own about maintainability, test coverage, or whether unusual patterns like an undetected daily API-call bug got caught before shipping.

Whether Anthropic or other labs publish a reproducible, clearly-defined measurement methodology, rather than a single headline percentage, would settle much of this debate; until then, expect the 90%-vs-50%-vs-metric-undefined argument to keep recurring each time an executive repeats the claim. Also worth tracking: whether GitHub's aggregate commit-volume data continues to accelerate, and whether independent security or maintainability audits of AI-heavy codebases surface more issues like the ones Tech Trenches found.