Google Builds Elite Team to Close the Coding Gap with Anthropic

Why It Matters

Enhancing Gemini’s coding skillset is critical for Google to stay competitive in the AI coding arms race and to lay groundwork for autonomous, self‑optimizing AI systems that could reshape software development.

Key Takeaways

• Google forms DeepMind “strike team” to boost Gemini coding abilities.
• Team targets long‑horizon tasks like building software from scratch.
• Sergey Brin urges models to become primary code developers.
• Internal tool “Jetski” usage metrics drive team performance.
• Training on Google’s private codebase may yield unreleasable but superior models.

Pulse Analysis

The race to dominate AI‑assisted coding has intensified as Anthropic’s tools outpace rivals in benchmark tests. Google’s response—creating a dedicated DeepMind team—signals a strategic shift from incremental model tweaks to a focused, long‑term effort on complex software generation. By targeting tasks that require reading multiple files, understanding user intent, and producing end‑to‑end solutions, Google hopes to transform Gemini from a helper into a primary developer, a capability that could unlock new enterprise offerings and tighten its grip on cloud services.

Internally, the initiative is more than a technical challenge; it reflects a broader vision championed by co‑founder Sergey Brin. In an internal memo, Brin linked advanced coding agents to the ultimate goal of self‑improving AI—systems that can rewrite their own code, run experiments, and iterate without human oversight. To accelerate this, every Gemini engineer must employ the Jetski agent for multi‑step tasks, and usage metrics are now tied to team performance. This mirrors Meta’s token‑consumption leaderboards, turning internal AI adoption into a competitive sport that drives rapid iteration.

The implications for the market are significant. If Google can harness its massive private code repository to train models that outperform public‑domain counterparts, it could deliver a proprietary edge in developer tools, cloud IDEs, and automated code review services. However, the reliance on unreleasable internal data raises questions about transparency and the ability to commercialize these advances. Competitors may respond by opening their own ecosystems or accelerating open‑source collaborations, keeping the AI coding battlefield dynamic and uncertain.