Google unveiled Gemini 3.5 Pro in May 2026, promising a release "next month". Two months later, the model is still not available to the general public. Google is even said to have restarted its entire pre-training from scratch, and is now targeting a release around 17 July. All this while several key researchers leave the company for competitors. An analysis of a series of setbacks that raises questions.
There is an unwritten rule in tech: when a company announces a product on stage with a date, it had better stick to it. Google is learning this the hard way in 2026 with Gemini 3.5 Pro, its most advanced AI model, whose release keeps slipping. Let's break down what's happening, without exaggerating or downplaying, because the story is telling about the state of the AI race.
A series of delays
The timeline speaks for itself. On 19 May 2026, at its I/O developer conference, Google presents the Gemini 3.5 family and its CEO Sundar Pichai announces that the Pro version will arrive "next month", i.e. in June. An audience of developers is said to have let out an audible sigh, given how accustomed Google's date promises have become to slipping.
June passes without a public release. Google then pushes back to July, citing quality adjustments following feedback from early enterprise testers. According to several sources, the company is said to have made a radical decision: abandon the previous base architecture and relaunch a full pre-training cycle, that long and costly phase we described in our article on AI training. The target now mentioned is 17 July, with no official guarantee.
Restarting a full pre-training is not a trivial decision: it means weeks of compute and colossal sums. If Google did it, it's because the issues found were deemed too deep to be patched at the surface. The concerns reported include excessive token consumption (and therefore costs too high for users), coding performance not yet up to the level expected of a flagship model, and reasoning on long tasks below the bar set. Better a delay than a botched launch: it's a quality choice, but one that costs dearly in terms of image.
The worrying signal: departures
A technical delay, in itself, is not dramatic. What makes this story more concerning for Google is the human context. In the same period, several senior researchers who worked directly on Gemini's architecture have left the company to join competitors, notably Anthropic.
Researchers of that calibre leaving just before the release of a major model is a strong signal. The reasons can be multiple: disagreements over technical direction, frustration with the pace of execution, or simply better opportunities elsewhere. Whatever the cause, the effect is twofold: Google loses expertise, and its competitors gain minds that know Gemini's inner workings intimately. The market has indeed reacted, punishing Alphabet's stock, Google's parent company, when these difficulties were announced.
What it says about the AI race
This affair goes beyond Google alone. It reveals that coding and agentic tasks have become the decisive battleground of AI in 2026. And it is precisely there that competitors like Anthropic, with Claude, have established themselves as the reference, and where Google struggles to close the gap despite its world-class research.
Revealingly, Google has chosen to release first its fast and economical version, Gemini 3.5 Flash, while holding back its flagship model. That is the exact opposite of the usual strategy of leading with your best model. This choice betrays a form of caution, or even difficulty in finalising the high end. That said, Google should not be written off: it has immense research power and resources, and its Flash version is already very solid. But the summit of AI has never been so contested.
What to take away
The delays of Gemini 3.5 Pro tell a story broader than that of a slipping schedule. They show that at a time of fierce competition, even a giant like Google can stumble, lose talent, and fall behind on the now-crucial ground of code and agents. In a race where everything accelerates, falling behind is costly, both in image and in talent.
For the observer, the lesson is clear: dominance in AI is guaranteed to no one, not even to the company that invented part of the foundational technologies. It remains to be seen whether Gemini 3.5 Pro, when it finally ships, will justify this long wait. If it lives up to its promises of a giant context and deep reasoning, the delay will soon be forgotten. If not, it will stand as the symbol of a moment when Google seemed, for the first time, to be running behind rather than ahead.