With the release of Gemini 3 last November, Google surged to the front of the pack of frontier AI labs, beating rivals on public benchmarks and pushing OpenAI to declare an internal emergency. Roughly nine months later, however, it is Google that finds itself under pressure to keep pace.
Google executives convened an all-hands meeting last week to address employee concerns over some high-profile leadership changes, including Demis Hassabis’ departure from day-to-day control. Meanwhile, co-founder Sergey Brin has spent recent months urging staff to focus on bringing Gemini back to the frontier of AI.
This serves as a tacit admission that rivals currently have Google beat.
Its user metrics paint a more positive picture, however. CEO Sundar Pichai announced yesterday that Google’s Gemini app had passed 1 billion monthly users. This makes it the 14th Google product to reach that milestone, and the fastest ever to do so. It is worth noting, however, that this figure includes anyone who opened the Gemini app, rather than paid users. Google has not disclosed its subscriber counts.
Scale is not the same as the frontier
The surprise change of role for Hassabis - founder of DeepMind and now Alphabet’s Chief Scientist - was announced by Pichai in an August 5 memo. CTO Koray Kavukcuoglu stepped in to take over daily operations, but the DeepMind CEO position remains vacant. The same day, former Chief Scientist Jeff Dean said he was leaving with three senior colleagues to start a company called Discovery Loop.
The restructuring that followed is what has elicited some concern from DeepMind employees. The UK-based AI lab, founded in 2010 and bought by Google in 2014, has maintained much of its independence and continued to operate from its London headquarters. Employees told Fortune that the recent changes could undermine that independence by shifting the balance of power toward Google’s Mountain View campus (a position rejected by Google).
TIME reported separately that this could extend to DeepMind’s safety organization. Chief AI Readiness Officer Lila Ibrahim will now report to senior Google executive James Manyika, while some safety teams previously under her could move under Google president of global affairs Kent Walker. Google maintains this will have no impact on the safety of its models.
Gemini 3.5 Pro has missed three deadlines
Google unveiled Gemini 3.5 Pro at its I/O developer conference in May. Initially, the company planned to ship it the following month. It has since missed three release dates, according to Fortune’s account of a difficult year inside the lab. A spokesperson for Google assured Forbes that the model is in testing and engineers have already begun working on future iterations.
Meanwhile, Google’s smaller, cheaper Flash models have filled the gap in the release schedule. Gemini 3.6 Flash is currently Google’s strongest publicly available model, but its performance leaves a lot to be desired. Micah Hill-Smith of AI benchmarking firm Artificial Analysis told Fortune that Flash ranks behind the leading models from every major U.S. AI lab, as well as two Chinese developers.
Gemini 3.6 Flash’s greatest weakness is coding - the most commercially significant function over the past year.
Competitor models are also increasingly able to execute complex, multi-step tasks without human oversight, a process known as agentic execution. If the first generation of AI models were in essence complex chatbots and search engines, these new functions turn them into digital assistants.
Brin identified Google’s shortcomings in this regard as far back as April. The Information reported that he had warned the firm’s AI engineers about the need to close this gap in an internal memo. With the flagship Gemini 3.5 Pro model still unreleased, it is impossible to gauge whether they’ve managed to do so yet.