Google Reports Results for Internal Coding AI GfG

Researchers from Google and other institutions published a paper in October describing a method for adapting large language models to companies’ own development data and reporting its practical effects. At Google, AI already creates or drafts more than 75% of code. However, general-purpose models trained on public data have struggled to work with the proprietary libraries, design patterns, and languages in internal repositories spanning billions of lines. The team developed Gemini for Google (GfG), an internal development version of Gemini.

Training used about 1 trillion tokens gathered from across the development process: completed code; code review comments and before-and-after changes; internal documents and Q&A; edits extracted from development interface histories that resolved build errors; and production improvements and operation logs. To prevent catastrophic forgetting (loss of previously learned capabilities during further training), the team combined an intermediate training intervention, which introduced internal data just before pretraining finished, with data replay, which mixed in general-purpose data for review. The researchers said this preserved general capabilities while reinforcing internal knowledge.

In a blind A/B test involving 29,000 internal developers, GfG reduced AI interactions needed to complete tasks by 23.1% and response time by 8.9% compared with Gemini 2.5 Pro. Code survival rate—the share of AI-generated lines that remained in the codebase without edits after 14 days—increased by 16.8%.

In practical use, an advertising team’s migration of 32-bit IDs to 64-bit integers took an estimated 50% less time than manual work, and AI wrote 80% of the code changes in the submitted change list. Google’s ECO system, which detects and optimizes inefficient code, saves computing resources equivalent to 500,000 CPU cores each quarter. Smart Paste, which adjusts copied code to fit its destination, accounts for more than 1% of all code newly written at Google. These figures describe usage and effects reported inside Google in the paper. The paper was published in October; the article’s November 13, 2026 publication date is not the paper’s release date. The 1 trillion tokens refers to training data scale, and the advertising migration’s 50% time reduction is an estimate.

Researchers from Google and other institutions published a paper in October describing a method for adapting large language models to companies’ own development data and reporting its practical effects. At Google, AI already creates or drafts more than 75% of code. However, general-purpose models trained on public data have struggled to work with the proprietary libraries, design patterns, and languages in internal repositories spanning billions of lines. The team developed Gemini for Google (GfG), an internal development version of Gemini.

Training used about 1 trillion tokens gathered from across the development process: completed code; code review comments and before-and-after changes; internal documents and Q&A; edits extracted from development interface histories that resolved build errors; and production improvements and operation logs. To prevent catastrophic forgetting (loss of previously learned capabilities during further training), the team combined an intermediate training intervention, which introduced internal data just before pretraining finished, with data replay, which mixed in general-purpose data for review. The researchers said this preserved general capabilities while reinforcing internal knowledge.

In a blind A/B test involving 29,000 internal developers, GfG reduced AI interactions needed to complete tasks by 23.1% and response time by 8.9% compared with Gemini 2.5 Pro. Code survival rate—the share of AI-generated lines that remained in the codebase without edits after 14 days—increased by 16.8%.

In practical use, an advertising team’s migration of 32-bit IDs to 64-bit integers took an estimated 50% less time than manual work, and AI wrote 80% of the code changes in the submitted change list. Google’s ECO system, which detects and optimizes inefficient code, saves computing resources equivalent to 500,000 CPU cores each quarter. Smart Paste, which adjusts copied code to fit its destination, accounts for more than 1% of all code newly written at Google. These figures describe usage and effects reported inside Google in the paper. The paper was published in October; the article’s November 13, 2026 publication date is not the paper’s release date. The 1 trillion tokens refers to training data scale, and the advertising migration’s 50% time reduction is an estimate.