Gemini 3.5 Pro Delay Continues

Google's highly anticipated Gemini 3.5 Pro is experiencing significant, repeated delays, sparking widespread concern about the company's competitive edge in the fast-paced AI race. Dubbed the "longest-awaited model of 2026," its launch has been pushed back multiple times. Reports cite persistent coding and reliability problems, senior researcher departures, and a possible complete retraining from its foundational pre-training phase due to a "structural problem." While Google focuses on enhancing other Gemini products and testing models with partners, the prolonged absence of 3.5 Pro is impacting market confidence and raising doubts about its eventual release. The tech giant faces increasing pressure amid intense industry competition.

It’s mid-August, and those who have been waiting a good chunk of the summer for the release of Gemini 3.5 Pro are getting restless.

Some are calling this the “longest-awaited model of 2026,” since a release deadline had previously been announced much closer to the beginning of the season. Apparently, other related models have been released, but Gemini 3.5 Pro is still locked away. Additionally, there are rumors of cancellation, with outlets like Geeky Gadget and others like Andrew.ooo contradicting each other about whether the Google Gemini page says 3.5 Pro is “coming soon.” (Take a look for yourself here).

“Google says it needs more time on coding and reliability. There’s also been senior-researcher churn. Rather than rush Pro, Google shipped the workhorse 3.6 Flash — described as improved coding/knowledge/multimodal at lower cost, using up to 65% fewer output tokens than 3.5 Flash.”

Widening the aperture, let’s look at some of the critical context of this delay, in pondering what’s next for the mega-firm amid a fast-paced and competitive LLM race.

In some internal Google reports, you can see that company reps suggest Gemini 3.5 Pro is “being tested with partners.”

That sounds a lot like what has happened at Anthropic, with Project Glasswing, and OpenAI, as powerful models get corralled before being publicly launched, just because they’re too powerful to be out there roaming around.

Google, you’d think, could make a plausible case for this: that given the sentiment in the market, and the propensity of agentic models to do unpredictable things, the leadership is just taking some extra time with Gemini, on principle. But according to what I’ve seen, nothing like that has been suggested.

Another reason for the lateness of Gemini 3.5 Pro might be that engineers are rolling out new Gemini App products instead.

“Yesterday, Google’s Josh Woodward asked for what fixes Gemini app users want,” reported Abner Li July 9 at 9to5Google. “The Gemini lead today identified the top 10 requests and Google’s progress.”

Among these were: more reliability for Google Workspace integrations, deep research improvements, the removal of watermarks from Nano Banana, and fixes for mobile app scrolling bugs. So maybe the Google engineers have just been busy.

Here’s another wrinkle: as this project experiences enigmatic delays, Google engineers continue to report that AI is increasingly doing the work. This has led to a debate between true-believers who insist on human coding, and those who are ok with letting the AI agent handle it.

If the back and forth has impacted project timelines, then it’s possible.

Some reports contend that the market is seeing troubling signs at Google, possibly relating to an inability to ship out models in a timely manner. For example, Wired reports on Jeff Dean and other prominent people pulling up stakes and heading out to places like OpenAI and Anthropic, direct competitors.

At the same time, those looking at Alphabet stock are seeing about a 12% drop since January, although there’s a lot of value built into the company over the past five years.

It’s reasonable, perhaps, to surmise that Google’s problems are related.

But this is the third deadline the company has blown for the model, and that leads some to conclude that the company is essentially going back to the drawing board.

“Pre-training is the initial, most expensive phase of building a frontier AI model,” writes Eloise Jones at Tech Times, “the run on a vast dataset that establishes a model's fundamental capability ceiling. Fine-tuning and reinforcement learning from human feedback can refine within those bounds; they cannot raise the ceiling. When Google reportedly chose to restart pre-training rather than continue refining, it was conceding that the original model's capability ceiling was in the wrong place, a structural problem, not a finishing problem.”

Basically speaking, when you delay a model release this much, it doesn’t spur confidence. Let’s hope that everything works out for Google, as we see more competition and fast moves on the market.