Engineers at Anthropic have shifted from writing code to reviewing it. CEO Dario Amodei has disclosed that Claude, the company's flagship AI model, now authors more than 80% of the code merged into Anthropic's production codebase, and many of the company's engineers no longer write significant code themselves.

Before Claude Code launched as a research preview in February 2025, that figure stood in the low single digits. In roughly 15 months, the tool went from research preview to the backbone of how one of the world's most prominent AI labs builds software.

The direction extends beyond one company, even if the magnitudes differ. Google CEO Sundar Pichai said in October 2024 that more than a quarter of new code at Google is generated by AI, and Microsoft CEO Satya Nadella said in April 2025 that up to 30% of code in some Microsoft projects is written by AI, though each company measures the share differently.

The numbers behind the shift

The productivity gains are striking. Anthropic reports an 8x increase in code shipped per engineer per quarter compared with its 2021 to 2025 baseline. Instead of spending hours writing functions, debugging edge cases, and optimizing performance, the company's developers now focus on high-level goals, system architecture, and oversight. Claude handles the grunt work — bug fixes, code optimization, and initial implementation — often autonomously.

According to internal assessments, the quality of Claude-generated code is now reportedly on par with what human engineers produce. Anthropic expects AI-authored code to surpass human performance within the next year.

The financial results reflect the productivity transformation. Anthropic has reported over 1,000% year-over-year growth in specific revenue metrics tied to Claude, with overall annual recurring revenue climbing by more than 500%.

Those numbers land in one of enterprise software's most competitive categories, where Claude Code contends with OpenAI's Codex agent and GitHub's Copilot, and where Anthropic's own codebase serves as a prominent showcase of the tool operating at scale.

How Claude Code changed the workflow

Before Claude Code arrived, AI assistance at Anthropic looked much like it does at most tech companies: autocomplete suggestions, occasional boilerplate generation, and perhaps a helpful code review comment.

Claude Code changed the equation. Rather than suggesting lines, the tool generates entire implementations that engineers review, adjust, and merge. Since launch, it has also moved from research preview to general availability, expanding in May 2025 beyond the command line to web, desktop, and sandboxed cloud environments.

The internal adoption curve was steep. Once engineers saw that Claude could handle not just simple tasks but complex bug fixes and optimizations with minimal human intervention, usage accelerated through 2025 and into 2026. Over that period, Claude's coding capabilities evolved from suggesting snippets in 2023 to operating as a full coding agent by 2025 and 2026.

The open questions from here are whether AI-authored code surpasses human performance within the year Anthropic expects, and how a self-reported figure from a single codebase translates to other organizations, where review practices, legacy systems, and risk tolerances vary.