Google Reworks Gemini AI Coding Team Amid Competition With Anthropic

Google is reshuffling an AI coding team it created only about two months ago, signaling that the company wants to strengthen Gemini’s programming capabilities as competition with Anthropic intensifies.

According to a report from The Information, Google DeepMind has expanded the responsibilities of its internal AI coding “strike team” to include a technique known as midtraining. The team was originally assembled to improve Gemini’s performance on complex, long-running software engineering tasks, but it will now also work on improving the underlying models before they reach their final training stage.

Google wants to improve Gemini before release

Midtraining is an intermediate phase that takes place after a large language model completes its initial broad training but before instruction tuning and task-specific optimization.

Instead of focusing solely on prompts, interfaces, or post-training refinements, developers can expose a model to carefully selected datasets that strengthen specific capabilities before the model is finalized.

For coding-focused AI systems, this approach could lead to more fundamental improvements in reasoning and software development tasks.

Previous AI research has suggested that midtraining can significantly boost performance in areas such as programming and mathematics, making it an attractive option for companies trying to improve model quality without dramatically increasing training costs.

The move suggests Google wants Gemini’s coding abilities to improve at the model level rather than relying primarily on external tools or prompt engineering.

Anthropic’s progress reportedly triggered the shift

DeepMind researchers believed Anthropic’s AI coding tools had moved ahead of Google’s offerings.

The report suggests that Google’s internal reorganization reflects increasing pressure to close that gap in one of the fastest-growing segments of generative AI.

No new Gemini product announced yet

Google has not announced whether the reorganized team will produce a new Gemini model or launch a separate developer-focused product.

The company has also not disclosed the team’s size, any performance targets, or a timeline for when the results of the new strategy might appear in public releases.

For now, the restructuring appears to represent an internal effort to strengthen Gemini’s competitiveness as AI companies continue racing to build the most capable coding assistants.

Meanwhile, the broader AI landscape continues to move quickly. Recent reports alleged that Alibaba attempted to distill capabilities from Anthropic’s Claude models, Anthropic recently introduced Claude Tag for Slack, and the newly released Mythos 5 AI reportedly faced restrictions following a warning from Amazon.

Via Neowin