Google's next flagship AI model is months late because it's falling short of Google's own bar

[Pixel Jay]

Google is running months behind schedule on delivering Gemini 3.5 Pro, its most powerful flagship model, according to Bloomberg, which cited 10 current and former employees describing the delay as a genuine source of frustration inside the company. The model was unveiled at Google I/O back in May with a targeted June general availability window, a deadline that has now come and gone with no public launch

The core issue, according to people familiar with the matter, is that the underlying technology isn't yet hitting Google's own internal bar, particularly on coding performance, prompting the company to take extra time trying to improve it rather than ship something that falls short. Other reporting on the delay points to token efficiency issues flagged by early testers and weaker than hoped long task, multi step reasoning, exactly the capabilities that power the AI coding agents currently driving a lot of enterprise demand, alongside a decision to scrap the existing Gemini 2.5 Pro base architecture entirely in favor of a full ground up rebuild on a native Gemini 3 foundation

Many Google engineers, researchers and managers are worried the company risks losing its edge as rivals Anthropic and OpenAI ship models that now exceed Gemini's capabilities in the meantime. Compounding the pressure, Google has reportedly lost several key researchers during this stretch, including a co-author of the influential T5 and Switch Transformer architectures, departures that raise real questions about the team's capacity to keep pace while simultaneously undertaking a full architectural rebuild

Google's own structural complexity isn't helping either, the company has to weave any new model across an unusually large product surface, search, maps, YouTube and more, with multiple layers of internal stakeholders needing to sign off before release, a coordination burden that smaller, more narrowly focused labs don't have to carry in the same way. The model is now reportedly targeted for a July 17 release, positioned as a cost effective alternative in the premium AI tier rather than an outright capability leader, though after this many delays the actual launch, and whether it can close the coding and reasoning gaps that triggered the rebuild in the first place, will say a lot about whether Google can still credibly compete at the very top of the frontier

Google Gemini Launch Delayed As Tech Falls Short Of Internal Goalswww.bloomberg.com

[HangmanPage_WCW]

Scrapping the entire base architecture and doing a full ground up rebuild instead of just patching the existing model is a bold, high risk call this late in the process, shows how serious the internal concern about falling behind actually is

[Voyager17]

Losing a co-author of T5 and the Switch Transformer during exactly this stretch is a rough coincidence of timing, that's real institutional knowledge walking out the door right when it's needed most

[StringTheory97]

The point about coding and long task reasoning being the specific weak spots is what actually worries me, those are exactly the capabilities enterprise customers are shopping for most aggressively right now

[Brooke_19]

Google's structural complexity, having to weave one model across search, maps and YouTube with layers of internal stakeholders, is such an underrated disadvantage compared to labs that only have to ship one product

[AlphaGareth16]

Positioning it as a cost effective alternative rather than an outright leader feels like a realistic repositioning given how badly this launch has slipped, better to be honest about the lane than overpromise again

[2026]

This is a good illustration of how much internal frustration and turnover can build up during a prolonged delay, the technical problem and the people problem seem to be feeding each other here

[ClaudioHerrera]

Feels like the classic big-company tradeoff. When your model has to plug into Search, Ads, Maps, YouTube, and Android, the bar is not just quality, it is consistency everywhere.

A smaller lab can ship something rougher and iterate fast. Google has to make sure it does not break core products.

That slows things down, but also explains why they hesitate to release anything half-baked.

[Attention Griffin]

Part of this might just be expectations catching up. "Flagship" now means beating everything else across reasoning, coding, multimodal, latency, and cost.

That is a huge checklist.

Missing on even one dimension can make the whole release feel underwhelming.

So delay starts to look like the safer option.

[William56]

The internal alignment point is underrated. Different teams want different things from the same model.

Search wants accuracy and citations, YouTube wants moderation and summarization, Ads wants targeting signals.

Trying to satisfy all of them with one system is messy.

No surprise timelines slip

[NightOwl94]

Better late than shipping something that damages trust.

We have seen what happens when models hallucinate in high-stakes contexts.

Google cannot afford that at scale inside Search.

So a delay might actually be a sign they are taking integration seriously.

[NeonSpectre11]

At the same time, the competitive pressure is real. If rivals keep improving while Google waits, perception shifts quickly.

People do not track internal complexity, they just see who is leading.

So there is a risk of being seen as lagging even if the final product is strong.

[Southern Jay]

This also highlights how different Google is from startups. A startup model can live in an API and evolve quietly.

Google ships into products used by billions.

That turns every model update into a product decision, not just a research milestone.

Totally different game

[EventHorizon25]

There is a chance the "falling short" narrative is more about internal benchmarks than real-world usefulness.

A model can be extremely capable and still fail some internal eval target.

Users might not even notice those gaps.

But internally, that is enough to delay.

[Aisha98]

Reminds me of how long it took Google to roll out major search changes historically.

They test endlessly because even a small regression at scale is massive.

AI just amplifies that caution.

More variables, more risk.

[Ella10]

End of the day, shipping late is only a problem if the result is still behind.

If they land something clearly competitive or better, the delay gets forgotten quickly.

Tech history is full of late but strong releases.

The real test is what shows up when it finally launches

[Tia91]

Some of this could also be about cost efficiency. A model that is powerful but too expensive to run at scale is not viable for Google.

So they might be optimizing inference costs as much as capability.

That kind of tuning can take time.

[CompiledVulture]

Delays like this often mean something fundamental is being reworked.

Could be data pipelines, alignment methods, or even architecture tweaks.

Those changes do not show up immediately but can pay off later.

So this might age well.

[Louise74]

There is also the talent angle. Coordinating large research teams is hard.

Different groups pushing different architectures or training strategies can slow convergence.

Not necessarily a bad thing, but it adds friction.

[William80]

Users tend to underestimate how much safety and policy layers add complexity.

It is not just about making the model smarter.

It is about making it behave correctly across millions of edge cases.

That is a huge engineering effort

[Robin13]

If Gemini 3.5 Pro ends up tightly integrated across products, it could feel more useful than a slightly stronger standalone model.

Context matters.

A "good enough" model in the right place beats a better one hidden behind an API.

[Kieron78]

The funniest part is how the narrative flips. A year ago people said Google was rushing to catch up.

Now the story is they are delaying because their standards are too high

Reality is probably somewhere in between.