Google Revamps New AI Coding Strike Team Amid Struggle to Catch Up With Anthropic
The Takeaway
- Google revamps AI coding strike team to catch Anthropic in lucrative market.
- Revamp comes as key AI researchers Noam Shazeer and John Jumper depart.
- The changes together show key challenges facing Google in AI compute allocation and coding.
Google is reorganizing its recently launched strike team working on AI coding tools to try to catch up with Anthropic in the most lucrative AI applications, according to people familiar with the changes.
The goal is for the months-old strike team to change the approach to training Google’s AI models to improve their abilities in both coding and other areas such as creating presentations, the people said. That aligns with efforts by both Anthropic and OpenAI to expand AI coding tools to other business functions. In addition to expanding the scope of the strike team’s work, the changes also formalize the structure of what was originally a short-term group.
The revamp coincides with two major executive departures that have fueled concern about Google’s ability to stay at the forefront of developing and adapting competitive AI models. While Google has broadly been one of the biggest winners of the AI boom, it has struggled with how best to allocate personnel and computing resources to compete with Anthropic and OpenAI.
Last week, star researcher Noam Shazeer quit abruptly to join OpenAI, as The Information was first to report. Shazeer and Google didn’t publicly provide a reason at the time, but people familiar with the situation said this week that his departure followed a change in his access to AI servers.
Shazeer’s departure was jarring to insiders and outside observers, after Google paid $2.7 billion in a licensing deal less than two years ago to bring him back following his departure in 2021 to found his own startup. A co-author of the 2017 paper on the so-called transformer architecture that has underpinned the current generative AI boom, Shazeer had recently been working on research to find new model architectures beyond transformers.
He told colleagues before leaving that the company had merged his compute allocation with that of another team, according to the people familiar with the situation. In an internal message announcing his departure, Koray Kavukcuoglu, chief technology officer of the DeepMind unit that leads Google’s AI efforts, said that Shazeer’s departure had happened quickly, one of the people said.
Then, on Saturday, John Jumper—a DeepMind vice president who was awarded a Nobel Prize along with DeepMind CEO Demis Hassabis for their work on protein structure prediction—announced he was heading to Anthropic. Jumper had recently been moved to the strike team, according to two people familiar with his work.
The two departures shook investor faith in Google, sending the share price of parent company Alphabet down 5% on Monday, its worst day in over a year. The stock has fallen another 1% over the past two days, though it remains up nearly 30% so far in 2026.
Google since last year has been one of the biggest beneficiaries of investor enthusiasm for AI, thanks to its strength throughout the AI value chain, from homegrown chips to leading frontier models to cloud and ad businesses booming because of AI advances.
Those assets have brought their own difficulties, though. Executives have struggled to manage internal tensions over competing demands for compute from an array of outside customers and partners—including Anthropic, a major Google Cloud customer—and internal efforts such as Shazeer’s project and the coding strike team.
That coding effort has been important enough that it drew direct involvement from both Kavukcuoglu and Google co-founder Sergey Brin. Coding has been the most lucrative part of the AI services market, and Anthropic’s dominance in the area has propelled it to an annualized revenue rate of $47 billion as of last month, more than three times the level in February.
OpenAI has been aggressively building its own coding tool, Codex, and both OpenAI and Anthropic have been expanding their tools to work on other white-collar applications.
The Google changes center on how AI models are trained. The work has broadly been divided into pretraining, or building the models, and post-training to adapt the models for specific uses. Google’s code strike team is now expanding into a midtraining team in between those stages, for training models on more specialized data, according to a person with knowledge of the changes.
The midtraining team will take some responsibilities for scaling model capabilities, including coding, that previously belonged to the post-training team, which will now focus on what it’s like for people to interact with the models. The goal is to separate those two areas to speed up capabilities improvement.
“While we are moving ahead with great speed, this is also a good time to structure the organization to have clear swimlanes and responsibilities,” Kavukcuoglu told employees in the email announcing the reorganization.
Coding has emerged as a major weakness in Google’s AI portfolio. It initially focused little on the area, in the belief that having the most capable base AI model would lead to coding abilities, according to a person involved in the effort.
Some developers criticized Google’s most recent model release, Gemini 3.5 Flash, for overly sycophantic answers and a higher price compared to previous Flash models. And developers met its Antigravity coding tool with mixed reviews after a buggy initial release.
The company has yet to release its next flagship model, Gemini 3.5 Pro, which it said at its May developer conference would arrive in June. At that time, it wasn’t expected to surpass Anthropic’s most advanced model, Mythos, according to a person who has tested it.