Gemini 3.5 Pro has been delayed for months: infighting within Google’s political factions leaves employees deeply frustrated

Google’s latest flagship model, Gemini 3.5 Pro, has been delayed—what’s exposed is not only technical bottlenecks, but also a prolonged internal political battle that’s wearing the company down. Bloomberg, citing 10 current and former employees, said the model has been held up for months because its coding capabilities have been unable to catch up with competitors, while internal infighting among factions within Google Cloud, DeepMind, and the Android teams has only made progress worse.
(Background: ReactBench benchmark for AI code-generation agents—GPT-5.6 Sol took the crown with 43.1%, as major language models step on Bug landmines)
(Additional context: Bloomberg says Anthropic’s IPO is slated to come as early as October, and the three major investment banks are arranging meetings between investors and executives)

Gemini 3.5 Pro, delayed by months, comes with Google’s internal anxiety that can’t be contained. Citing 10 current and former Google employees, Bloomberg reported that this flagship model—highly anticipated and designed to take on Anthropic and OpenAI—has fallen behind its original delivery timeline by several months, and the key sticking point is that its coding abilities have still not been able to catch up.

The market originally expected 3.5 Pro to debut at this year’s May I/O developer conference, when CEO Sundar Pichai even talked it through to June. Now that July is already halfway over, the hand-in date remains unknown. Engineers, AI researchers, and management alike are permeated with frustration as they watch their rivals’ models gain ground while their own ace is still stuck in refinement.

Factional infighting is more difficult to manage than the competition

Google’s product lines are vast. To weave AI models into scenarios like search, maps, and YouTube, releases have to pass layers of checks by key stakeholders—this is one of the structural reasons for delays. But Bloomberg points to an even more troublesome root cause: internal factions competing with each other, dragging everyone’s pace.

Co-founder Sergey Brin has repeatedly argued that Google should move faster on AI coding. In practice, however, Google Cloud, Google DeepMind, and the Android teams operate independently. Meanwhile, they’re simultaneously building AI coding tools for developers, and even the consumer product teams have gotten involved.

One side is a multi-headed operation; the other is internal “purists” pushing back. Some engineers believe that truly important code should be written by humans to meet Google’s standards. In the rollout’s early phase, employees were even restricted from using Gemini to write or analyze code, on the grounds that proprietary code might leak into AI training data. That policy was only loosened later.

Chief AI architect Koray Kavukcuoglu is currently working with major engineering teams to unify internal tools. DeepMind also earlier this year formed an AI coding team led by research engineer Sebastian Borgeaud, trying to clean up internal chaos. But the reality is that when engineers are tasked with generating code using AI, they often get stuck due to internal compute shortages and resource contention hitting capacity limits. Some researchers who are unhappy with the company’s position in the AI race have already turned to top labs like Anthropic.

The cost of boiling an entire ocean

A former Google employee described the difficulty of getting leadership in every department to move in the same direction as “like trying to boil the whole ocean.”

When ChatGPT burst onto the scene at the end of 2022, Google announced “code red” in an effort to break through bureaucracy and internal competition. Now sprinting down the AI track has long become routine for the company—it’s just that the direction of the sprint hasn’t yet converged into a single line.

Google has plenty of cards. Its most-used products worldwide are the main entry point for most people to access generative AI, and these interactions themselves can continuously generate data that helps make answers smarter. Add its multimodal capabilities for processing inputs like images and video, along with progress on “AI world models” that can simulate the physical world—those are moats that Anthropic and OpenAI can’t match, at least for now.

But the moat can hold in the present schedule—it can’t hold against the timetable: when competitors roll out strong models one after another, Google still hasn’t even settled the internal debate over which coding tools it should use. In this generative AI game, the late giant still looks like it’s struggling to keep up.