Google Says 75% of Its New Code Is Now Written by AI — Up From 25% Just 18 Months Ago

Google CEO Sundar Pichai disclosed at Cloud Next '26 that 75 percent of new code written at Google is now generated by AI and reviewed by human developers. The trajectory of that number is as significant as the number itself: in October 2024, the figure was 25 percent. By fall 2025, it had reached 50 percent. The doubling in the eighteen months since October 2024 suggests an acceleration rather than a plateau — AI code generation is not settling into a steady-state contribution; it is still expanding its share of Google's development output. The engineers reviewing that AI-generated code are increasingly being positioned not as writers who receive AI suggestions, but as reviewers who validate AI output — a subtle but structurally significant role shift.

The Migration Case Study

The most concrete evidence Pichai offered for the productivity implications of this shift was a complex internal code migration that the company completed six times faster than a comparable migration completed a year earlier. The migration involved AI agents and human developers working collaboratively — the agents handling the mechanical transformation work (identifying call sites, applying refactoring patterns, generating tests) while developers provided architectural judgment and reviewed output. The 6x speed improvement is a striking data point, but it comes with important context: migration tasks are among the most favorable for AI assistance because they involve applying consistent transformations at scale across a large codebase, rather than original design or architectural reasoning, where AI performance is less reliable. Whether similar multipliers apply to greenfield development or complex feature work is a more open question.

Gemini Dominates Internally — But Not Exclusively

Google's engineers primarily use the company's own Gemini models for AI-assisted development. However, a report from Business Insider noted that some Google DeepMind employees are permitted to use Anthropic's Claude Code for coding tasks — an acknowledgment that Google has not yet built a tool that fully matches Claude Code's developer experience for certain workflows. According to The Decoder's reporting, a dedicated team within Google DeepMind is now working to close this gap. The internal use of a competitor's tool by a team inside the company's own AI research division is a rare admission of a competitive shortcoming, and suggests that the developer tooling race between Anthropic and Google is still genuinely contested — even at Google itself.

What Agentic Workflows Mean at Scale

Pichai's framing of the shift as a move toward "agentic workflows" points to where Google sees the next phase of AI-assisted development. Rather than AI that completes individual lines or functions on request, agentic workflows involve AI systems that operate with increasing autonomy across multi-step tasks — writing a function, writing its tests, running them, debugging failures, and submitting a pull request without requiring human input at each step. At Google's scale, where thousands of engineers contribute to an enormous shared codebase, the efficiency implications of AI agents handling routine development tasks compound dramatically. The 75 percent figure suggests Google has already internalized this shift — the question for the rest of the industry is how quickly the same transition occurs at organizations with less AI infrastructure and fewer proprietary models to deploy.