Google's new strike team wants AI to train AI

When your own elite AI researchers prefer your rival's tool to write their code, you don't just hold a meeting—you call in the co-founder.

According to a report from The Information, Google co-founder Sergey Brin is personally rallying DeepMind to out-code Anthropic. He has set up a dedicated internal strike team tasked with closing the gap between Gemini and Claude.

Forced dogfooding and an internal leaderboard

The new strike team is led by research engineer Sebastian Borgeaud, who previously oversaw pretraining at DeepMind. He is working directly under CTO Koray Kavukcuoglu and Brin himself.

The urgency is largely driven by brutal internal feedback: DeepMind's own researchers reportedly rate Anthropic's Claude as superior to Gemini for day-to-day code generation.

The automation divide between the two companies is also stark. While AI currently generates about 50% of Google's code, Anthropic has reportedly pushed its internal coding toward near-total automation.

To bridge that gap, Google is turning up the pressure inside its own walls. Gemini engineers are now required to deploy internal agent tools on complex workflows, with adoption openly tracked on a company leaderboard called Jetski. Beyond forced usage, the squad is focusing on long-horizon coding tasks and training models directly on Google's proprietary codebase.

The real prize: AI that builds AI

While matching Claude is the immediate trigger, Brin's vision goes way beyond shipping a better Copilot alternative.

In an internal memo, Brin told staff that the real prize is building an AI system capable of training the next AI—and coding capability is the shortest, most reliable route to get Gemini there.

Code provides unambiguous feedback: it compiles or it breaks, it passes tests or it fails. If a model can reliably architect code and automate engineering loops, it effectively unlocks self-improving recursive development.

Why it matters

This is not a cosmetic marketing pivot; it is Google re-architecting its own internal engineering engine to survive an escalating talent and efficiency war.

For external developers, the fruits of this strike team won't show up in the Gemini API overnight. But Google's pivot toward long-horizon tasks means future updates will likely target sprawling, multi-file codebases rather than simple single-function completions. If Google succeeds, developers might finally get an agent capable of handling real refactors rather than trivial autocomplete.

Across the industry, Anthropic has built an impressive moat purely through developer loyalty. After dominating headlines at the end of 2025, Google entered early 2026 looking somewhat sluggish on the coding front. Brin stepping onto the front line proves that leadership sees code generation not as a feature, but as the foundational bedrock of AGI research.

We think Google's strategy here is surprisingly pragmatic. Instead of chasing synthetic benchmark scores, they are weaponizing their own proprietary codebase and using the Jetski leaderboard to force internal dogfooding on complex, messy internal tasks. The first lab whose AI can autonomously build and train the next generation model will run away with the race—and Google wants to make sure Anthropic doesn't get there first.