Google DeepMind has assembled a strike team of researchers and engineers with one urgent mission: supercharge its coding models. The move is a direct response to feeling the heat from rival Anthropic.
Led by DeepMind research engineer Sebastian Borgeaud, the team also has direct involvement from Google co-founder Sergey Brin and DeepMind CTO Koray Kavukcuoglu. The catalyst? DeepMind researchers believe Anthropic's coding tools have already surpassed Gemini's code-generation capabilities.
In a recent memo, Brin wrote that the team must "urgently close the gap in agent execution capabilities and turn the model into a primary developer for writing code" to win the final sprint. (Here, 'agent' refers to AI that can handle multi-step tasks.)
The gap isn't just theoretical—it's quantifiable. In January, Anthropic's Claude Code lead Boris Cherny claimed the company writes "almost 100%" of its code with AI. In contrast, Google CFO Anat Ashkenazi said during a February earnings call that coding agents handle only about 50% of code work internally.
The Hard Problems
The strike team is focusing on long-horizon coding tasks, like building new software from scratch. These jobs require a model to read across multiple files and understand user intent—areas where current AI coding assistants struggle the most.
Training data is getting an overhaul, too. Google is now using its own private code repositories to train models, as the gap between internal and public code is significant. General-purpose coding models simply don't perform well enough on Google's internal projects. These internal-only models can't be released publicly, but they can help iterate on the versions that eventually are.
Internal Push
To drive adoption, Google has created an internal leaderboard called Jetski that ranks teams on their use of coding tools. Teams outside DeepMind are starting to mandate AI training. Brin's memo also reportedly requires every Gemini engineer to use the internal agent for complex, multi-step tasks.
The Long Game: Self-Improving AI
The ultimate goal, as Brin calls it, is "AI takeoff"—self-improving artificial intelligence. He's repeatedly told employees that boosting coding capability is the key to this step. Paired with AI that can do math and run experiments, it could theoretically automate the work of AI researchers and engineers at scale. OpenAI already uses similar internal tools to help its researchers generate experimental code faster.
This isn't the only front in the AI coding war. Competitors like Anthropic are pushing forward with models like Claude Opus 4.7, which boasts smarter coding and better multi-step workflow management. Meanwhile, Elon Musk's xAI is preparing its own entry into the arena with Grok Build and Grok CLI, aiming to challenge the current ecosystem dominated by Claude Code, OpenAI's Codex, and Google's own tools.