Gemini 3.5 Pro was announced on 19 May 2026 for a June release. It is now mid-August, it is still not publicly available, no date has been confirmed, and some observers are now suggesting it may be scrapped in favour of a future version. Meanwhile, Gemini has crossed one billion monthly users. These two facts tell the same story.
We covered this story twice in July, then set it aside as it continued to develop. That was a tracking error, and the topic deserves a full revisit, because it has become far more than a scheduling story.
The timeline, which is starting to drag
On 19 May, at the Google I/O conference, Sundar Pichai announced the Gemini 3.5 family and said the Pro version would arrive the following month. Part of the audience let out a sigh, as Google's date promises have already developed a habit of slipping.
June passed. Then July, with a widely reported date of the 17th, which we covered in our July article. That date passed too. It is now mid-August, and the situation has not changed: no gemini-3.5-pro identifier in the public documentation, no spec sheet, no pricing grid. Only Gemini 3.5 Flash has shipped, along with the 3.6 Flash generation we covered in late July.
One telling detail: Google has apparently registered additional Flash model names, which some read as fallback options. The hypothesis now circulating is that the Pro model could be quietly dropped in favour of a future generation.
What is blocking things, based on available evidence
The reported reasons converge on three points, none of which Google has officially confirmed.
A structural problem. DeepMind reportedly abandoned its initial base model after finding deep flaws, particularly in recursive tool calling and vector graphics generation. Restarting full pre-training means weeks of compute and considerable expense: you do not do that for a minor defect.
Reliability. Reports cite hallucination rates and consistency in real-world conditions falling short of Google's internal targets.
Departures. Several senior researchers who worked on the architecture have left the company, something we already noted in our first article and which appears to have continued.
While the flagship remains stuck, Gemini crossed one billion monthly active users on 11 August. This is not contradictory, it is revealing: distribution matters more than the model. Google is placing Gemini in Search, in Android, in its office tools, and soon in Siri. A user who encounters the AI because it is already there does not compare benchmarks. This is exactly the issue we described regarding opening Android to competing assistants: the interface battle is won independently of the model battle.
What this story really says
Three lessons emerge, and they go beyond Google's case.
Restarting a full training run has become conceivable. Two years ago, such a delay would have been unthinkable for a player of this size. That Google is accepting it signals that quality now comes before schedule, at least in this specific case. One could see this as a form of seriousness.
Date-based communication is a trap. Announcing an unfinished product on stage creates a credibility debt that is then paid on every subsequent announcement. This is probably the most costly lesson of this episode.
Delay costs you talent. A researcher who sees their work postponed indefinitely, in a market where competitors ship every month, has objective reasons to leave. And each departure slows things down a little more.
What to take away
The real question is no longer when Gemini 3.5 Pro will ship, but whether it will ship under that name. A model announced in May, delayed four times, arriving in September against competition that has shipped three generations in the meantime, would find itself having to prove an enormous amount.
It is possible that the most rational decision is to skip this step and focus efforts on the next generation. That would make for an unpleasant headline for forty-eight hours, and it would probably be better than shipping a model that disappoints after such a delay.
For a developer, the advice remains what we gave in July: build on what exists and works, keep an abstraction layer to switch models, and build nothing around an unconfirmed date. That advice has aged well.