Highlights:
- Google DeepMind has formed a specialized **strike team** led by engineer **Sebastian Borgeaud** to rapidly improve **Gemini** models on complex coding tasks.
- The team is training on Google’s massive **2 billion-line** internal codebase to close the gap with **Anthropic’s Claude**, which reportedly writes nearly 100% of new code at Anthropic compared to about 50% at Google.
- Sergey Brin and DeepMind CTO are directly involved, pushing for **agentic AI** that can handle multi-step engineering challenges and eventually self-improve.
- Internal assessments show **Claude** outperforming Gemini in coding, creating urgency and some internal tension among staff.
- This move signals Google’s aggressive push toward AI systems that can act as primary developers, potentially transforming how software is built across the industry.
I recently came across some interesting internal developments at Google DeepMind that caught my attention. The company has quietly assembled a strike team to boost its Gemini models’ coding abilities, and I find this quite telling about the current state of the AI race.
Sebastian Borgeaud, a seasoned research engineer who previously led pre-training efforts for Gemini, is heading this dedicated group. Their main mission is clear: train the models on Google’s enormous 2 billion-line private codebase to tackle complex, real-world engineering problems that current versions still struggle with.
When I dug deeper, I realized why this push feels so urgent. Internal evaluations apparently show that Anthropic’s Claude is significantly ahead in coding performance. At Anthropic, Claude is said to write nearly 100% of new code, while at Google the figure hovers around 50%. That gap is substantial, and it prompted Sergey Brin to get personally involved through an internal memo.
Brin is advocating for a strong shift toward agentic AI — systems capable of handling multi-step tasks autonomously, reading large codebases, understanding user intent, and even writing complete software from scratch. I see this as more than just catching up; it’s Google aiming for self-improving AI that could one day automate large parts of its own research and development.
Here’s a quick comparison of the current coding landscape as I understand it:
- Anthropic (Claude): Writes ~100% of new code internally
- Google (Gemini): Writes ~50% of new code internally
- Goal of DeepMind strike team: Bridge this gap and move toward full agentic coding capabilities
The effort also highlights some internal divides. A few DeepMind researchers reportedly prefer using Claude for their own work and have raised concerns about access and priorities. Still, the company is doubling down, with both Brin and DeepMind’s CTO directly overseeing progress.
I believe this development matters a lot for developers and tech companies. If Gemini catches up or surpasses Claude in coding, we could see faster software development cycles, fewer bugs in AI-generated code, and more reliable autonomous agents that handle entire projects with minimal human supervision.
In my view, this strike team represents Google’s recognition that coding prowess is now a critical frontier in the race toward more advanced AI. It’s not just about writing better code — it’s about building systems that can evolve and improve themselves.
My personal take
I think we are entering a phase where AI coding agents will become the primary developers in many organizations. Professionals working in software engineering should start experimenting with the best available tools today, whether it’s Gemini, Claude, or others, to stay ahead. Companies need to prepare their workflows for deeper AI integration, including better review processes for AI-generated code and updated security practices.
This move by DeepMind will likely accelerate innovation across the industry. In the coming months, I expect to see noticeable improvements in Gemini’s ability to handle long-horizon coding tasks. For all of us who build with AI, that means more powerful assistants and potentially faster progress toward truly intelligent systems. The competition is heating up, and that’s ultimately good news for everyone working in tech.
FAQs
What is the DeepMind strike team focused on?
The team, led by Sebastian Borgeaud, is working to enhance Gemini models’ performance on complex coding tasks using Google’s massive internal codebase, aiming to build more capable agentic AI systems.
Why is Google pushing so hard on coding AI now?
Internal assessments revealed that Anthropic’s Claude significantly outperforms Gemini in coding, writing nearly 100% of new code at Anthropic compared to about 50% at Google, prompting urgent action from leadership including Sergey Brin.
What does “agentic AI” mean in this context?
Agentic AI refers to systems that can autonomously handle multi-step tasks, understand user intent, navigate large codebases, and act as primary developers with minimal human intervention.
Will this affect developers using Gemini?
Yes, successful improvements should lead to stronger coding assistance in Gemini, making it more competitive with Claude for real-world software engineering and long-horizon projects.
What should software teams do in response?
Teams should test current Gemini and Claude coding capabilities, update code review processes for AI-generated output, and prepare workflows for greater AI autonomy in development.