- 75% of Google’s new code is generated by AI and reviewed by human engineers, up from 25% in October 2024 and 50% in fall 2025.
- CEO Sundar Pichai disclosed the figure in a blog post this week, citing a shift toward “agentic workflows” with increasing AI autonomy.
- An internal code migration completed with AI assistance ran six times faster than a comparable project from the prior year, according to Google.
- Some Google DeepMind engineers currently use Anthropic’s Claude Code, per Business Insider, while an internal team works to build a competing tool.
What Happened
Three-quarters of all new code written at Google is now generated by AI before being reviewed by human engineers, CEO Sundar Pichai disclosed in a blog post this week, as reported by The Decoder. The figure has climbed steadily: AI-generated code stood at 25 percent in October 2024, rose to 50 percent by fall 2025, and reached 75 percent as of April 2026.
Pichai stated that Google is now shifting toward what he described as “agentic workflows,” in which AI systems take autonomous action rather than responding only to individual developer prompts. The company’s engineers use Google’s own Gemini models for these tasks.
Why It Matters
Google’s disclosure provides one of the most detailed continuous public data series on AI coding adoption at a major technology company — three benchmark figures spanning 18 months. By comparison, most technology peers have released narrower metrics: Microsoft reports GitHub Copilot code-acceptance rates for individual suggestions, and Meta has described AI assistance in specific workflows without providing an equivalent organization-wide figure for new-code share.
The 18-month progression from 25 to 75 percent is notable for its pace. Industry analysts covering developer productivity tools have generally projected slower adoption curves for AI code generation at enterprises of Google’s scale, given the complexity of internal codebases and review processes. The disclosure establishes a public benchmark that competitors and researchers will now use as a reference point.
Technical Details
Gemini models handle the bulk of Google’s AI code generation workload. Pichai’s blog post cited a specific performance outcome: an internal code migration completed with AI-agent assistance ran six times faster than a comparable migration performed one year earlier. The company attributed this result to collaboration between AI agents and human developers working in tandem rather than either operating independently.
The 75 percent figure refers to new code initially generated by AI before human review — it does not describe code review acceptance rates, define how “new code” is counted across different project types, or indicate the cumulative share of Google’s total codebase that is AI-authored. The 6x speedup claim was published in Pichai’s blog post without an independently verifiable methodology or detailed comparison criteria.
Who’s Affected
Google’s software engineering workforce is the primary user base for these tools. Despite this, Business Insider reported that some engineers at Google DeepMind are currently permitted to use Anthropic’s Claude Code — a direct competitor’s product — for internal tasks. The report attributed this to gaps in Google’s internal developer tooling, indicating that Gemini-based coding tools do not yet fully satisfy the needs of the company’s AI research engineers.
A dedicated team within Google DeepMind has been assembled to build an internal alternative to Claude Code, per Business Insider. That a portion of Google’s core AI research workforce relies on a competitor’s coding tool is a concrete indicator of where the internal tooling gap currently sits.
What’s Next
Pichai indicated that agentic AI workflows will continue expanding across Google’s engineering functions, suggesting the 75 percent figure represents a current milestone rather than a ceiling. The Google DeepMind effort to build an internal Claude Code alternative is the most concrete near-term development to track.
Pichai did not specify a timeline for further increases in AI-generated code ratios or detail the architecture of the expanded agentic workflows the company plans to roll out across its engineering organization.
