Google's AI division gets Claude while the rest of the company eats its own dogfood, and the performance reviews just got interesting.
The Summary
- Google DeepMind employees now have access to Anthropic's Claude for coding, while most Google engineers are restricted to internal Gemini models.
- Engineers outside DeepMind report Google's internal coding tools lag behind Claude, creating a two-tier system as AI usage becomes tied to performance reviews.
- Google is mandating AI adoption across teams with specific goals that affect career progression, while simultaneously limiting which tools most engineers can use.
The Signal
Google just created a natural experiment in agent-assisted development, and the independent variable is access to the best tool for the job. Some DeepMind employees can use Claude for coding. Everyone else is stuck with Gemini. The results matter because Google is now tying AI usage to performance reviews.
This is not about hurt feelings. This is about velocity. When you tell engineers their job performance depends on using AI tools, then give them inferior tools, you create a measurement problem. Are lower-performing teams actually worse at AI adoption, or did they just get worse tools? Google won't know.
"When performance reviews depend on AI usage, tool quality becomes a career variable, not just a productivity metric."
The dogfooding defense makes sense for consumer products. Use Gmail to make Gmail better. But coding agents are different. They are productivity multipliers for the people building everything else. If your DeepMind team codes 30% faster because Claude handles boilerplate better, that gap compounds daily. The resentment compounds too.
Here's what Google is actually testing:
- Can you mandate agent adoption without providing best-in-class agents?
- Will engineers find workarounds (they will)
- Does restricting tools hurt more than the IP security risk of external models?
Meta apparently decided external AI models are fine. Google DeepMind decided Claude is fine for its own people. But the rest of Google gets the internal version. That split tells you something about confidence levels in Gemini's coding capabilities.
The performance review angle changes everything. Google is not just suggesting AI usage. It is requiring measurable AI goals that affect compensation and promotion. You are being graded on how well you use tools you did not choose and may not trust. If those tools are objectively worse than what the market leader uses, you are being graded on a curve that does not account for equipment quality.
The Implication
If you are building agent-first companies, watch this closely. Mandatory adoption without best-in-class tools is a recipe for shadow IT and workarounds. Your engineers will find Claude, or Cursor, or whatever actually works. They will just hide it.
The smarter play: admit that agent tools are infrastructure, not products to dogfood. Let your teams use what makes them fastest, then figure out why your internal tools are not winning. Google has the talent and resources to build the best coding agents in the world. The fact that DeepMind chose Claude anyway is the signal.