Google is reorganizing the AI coding strike team it stood up only a couple of months ago, widening its mandate to include model training as it tries to close the gap with Anthropic. The Information reported the move on June 25, and the timing is rough: two researchers central to the work it was supposed to fix are heading out the door, both to Anthropic.
The team formed in April 2026. That alone was a tell. You do not assemble a strike team for a problem you think is going fine.
What the team is now chasing
Originally the group focused on coding tools and agents, the product layer. According to Neowin, Google DeepMind is now stretching its scope into midtraining, the stage that sits between a model's broad initial pretraining and the post-training that teaches it to follow instructions. The idea is to feed Gemini better, more structured data before the final polish, rather than papering over weak coding ability with smarter prompts and interfaces.
There is research backing the bet. Midtraining tends to pay off most for code and math, the kinds of tasks where a model has to graduate from general language fluency to something more rigorous. So the direction makes sense on paper. Whether Google can execute it while bleeding the people who would do the executing is the open question.
The departures are the actual story
Jonas Adler, who led AI coding work inside DeepMind, and Alexander Pritzel, who worked on pretraining, are both leaving for Anthropic. Bloomberg broke the news on June 24, citing people who spoke anonymously because the moves had not been announced. TechCrunch confirmed the same day.
Read that lineup again. One worked on coding, one worked on training. Those are precisely the two layers the reorganization is meant to repair. They are not the only losses either. Nobel laureate John Jumper left for Anthropic the week before, and Gemini co-lead Noam Shazeer is going to OpenAI. Adler and Pritzel had both worked with Jumper on AlphaFold, so Anthropic is reassembling a cluster, not just poaching individuals.
"There's a lot of talent movement between all the leading labs and we win our fair share of the top talent," said DeepMind CEO Demis Hassabis at an event in Cannes, which is the sort of thing you say when you have just lost four people in a week.
Why the exodus? Money is part of it, but the sharper pull is timing. Anthropic and OpenAI are both circling IPOs, and a pre-IPO equity package at a fast-growing lab beats a comfortable salary at a company already worth trillions. There were also fights over compute. Before Shazeer left, capacity dedicated to one of his projects got reassigned to a London team, per Bloomberg's reporting.
The number Google does not love
Google CFO Anat Ashkenazi has acknowledged AI writes roughly 50 percent of Google's code. Anthropic has said publicly that Claude writes close to 100 percent of its own. Treat both figures with some suspicion, because "AI writes the code" means wildly different things depending on who is counting and what counts as a line. Still, the gap is the kind of thing executives notice, and co-founder Sergey Brin reportedly put it bluntly in an internal memo about turning Google's models into primary developers of final code.
The product backdrop does not help the optics. Google's Antigravity coding suite has shipped and works, but its pricing and credit system have drawn repeated complaints, and the default Gemini 3.5 Flash model gets expensive fast in real agentic use once you count how many turns it burns through. The flagship Gemini 3.5 Pro has not shipped yet. Reviewers keep describing it as the release that might fix the coding and cost gaps, which is another way of saying the current lineup has not.
So does the reorg matter
Maybe. A structural fix only works if the people inside it stay, and right now the direction of travel points one way. One complication worth flagging: reports indicate the strike team trains on Google's proprietary codebase, which would make any direct public release of those gains messy.
No release date, team size, or performance target has been disclosed. The next real signal is Gemini 3.5 Pro. If it lands and the coding numbers jump, the midtraining pivot looks vindicated. If it slips again, the talent leaving for Anthropic will look less like individual choices and more like a verdict.
