Google DeepMind has created a dedicated group of researchers and engineers tasked with upgrading its AI coding models, according to reporting from The Information. Internally, some DeepMind staff reportedly see Anthropic’s coding assistants as stronger than Google’s own Gemini-based tools, which has sharpened the urgency. Senior Google leadership, including the company’s co-founder and its chief technology officer, are said to be directly involved in steering the effort.
The new team is prioritising models that can generate production-ready code for Google’s internal use, which means training on the company’s proprietary codebase. Models trained on that sensitive internal code cannot be shipped directly to consumers, but they can act as a foundation to improve versions that are safe to release publicly. Leadership of the strike team sits with a senior Google DeepMind research engineer who previously oversaw pretraining work for the organisation, according to multiple sources cited by the publication.
Engineers on the team are reportedly focused on boosting performance for long-running, complex coding tasks rather than simple code snippets. That includes making models better at navigating large code repositories, reading across many files and inferring what a developer is trying to achieve from relatively sparse instructions. The aim is to build tools that can handle real-world software engineering workflows, not just demo-friendly prompts. Those capabilities matter most inside a company with sprawling, decades-old code.
Competitive pressure comes from Anthropic, which has publicly framed AI coding agents as central to its own development process. The company has said its tools now handle the bulk of engineering work, with its Claude Code product used for nearly all code generation. Google’s finance chief has separately disclosed that around half of Google’s code is already being produced by internal coding agents, a figure that shows how quickly AI is reshaping software development inside major tech firms.
