Our chart this week shows new code at Google written by AI jumping from 25% in October 2024 to 75% today, disclosed by Sundar Pichai at Cloud Next in Las Vegas this week. Four data points, each roughly doubling the last. The more useful number buried in Pichai’s remarks is six: a recent internal migration finished six times faster than the same job a year ago.
Anthropic was also at Cloud Next in force. Claude runs on Vertex, and Claude Code is now reportedly on a $2.5 billion annual run-rate. Customers are increasingly mixing OpenAI, Claude and less so Gemini agents inside the same workflow, including, awkwardly for Google, inside Google itself.
Which brings us to Steve Yegge. Over the past fortnight, the former Google engineer has posted a series of claims, sourced to current Googlers across multiple orgs, that paint a rather different picture from Pichai’s victory lap. Yegge alleges a two-tier system inside Google: DeepMind engineers use Claude Code daily, while most of the rest of the company is pushed onto internal Gemini variants that, in his sources’ view, are not yet as effective for agentic coding. When someone internally proposed equalising access by removing Claude for everyone, DeepMind reportedly pushed back so hard that several engineers threatened to leave.
Google’s leadership have disputed this in forthright terms. Demis Hassabis called the original post “absolute nonsense”. Addy Osmani, a director at Google Cloud, said more than 40,000 Google engineers now use agentic coding weekly, and Paige Bailey at DeepMind noted teams with agents “running 24/7”.