Gemini 4 Leaks: What They Show, What Is Unverified, and Why Google's Data Matters

Important up front: none of what follows is confirmed by Google. Everything here is leaked, rumoured or posted by third parties. Treat it as speculation, not news.

Why I made this video

Most "Gemini 4 leaked" videos are polished and scripted. I wanted to do the opposite: open a browser, look at the leaks as a normal user would, and react live. No script, no strategy.

Before we start, one question to keep in mind. Every AI model needs data to produce a result. Who has the most data right now? Google. That, more than any benchmark, is the angle that could change things. Google either makes a serious comeback with workflows and automation built on that data, or this becomes another flop. Either outcome would surprise nobody.

Step one: ask Gemini itself

I asked Gemini directly whether the Gemini 4 leaks are real. The short answer it gave: no, they are not officially confirmed, and parts of them appear fabricated or unverified. That sets the tone. From here on we are speculating.

The "Argon" checkpoint claims

One of the first accounts to post about it on X shared what it described as the first output from a Gemini 4 Pro checkpoint, internally code-named Argon: a block-built 3D scene that looks like something you would make in Blender. The claims attached to it:

• The output took about 2.4 minutes on high thinking effort.

• An output limit of around 256K tokens, compared with 64K in previous Gemini models.

• A context window of 2 million tokens at launch, roughly double what competitors offer today.

If those numbers hold, more context means much bigger jobs in one go: whole codebases, long document sets, full projects in a single request.

Benchmarks and pricing

Several posts claim wild benchmark scores, high speed and low pricing. The figures circulating are roughly $2.25 per million input tokens and $11.25 per million output tokens. Nobody has verified the benchmarks or the prices. If the pricing is real, it would put pressure on the whole market.

The 3D demos

This is where the leaks get interesting. The examples shared on Reddit and X include:

• A detailed race car

• A 3D city building, so architecture as well as objects

• A ship, and transformations of scenes

• The classic "pelican on a bicycle" test, with customisation

• A pagoda shown side by side: the older Flash model's version next to the claimed Gemini 4 Pro version. The difference is night and day.

If any of these came from one prompt with no outside software, that is a big shift. Anyone who has tried to get results like these in Blender or another 3D tool knows how much work they take.

What it would mean if it is real

• Websites will not stay the same. Interactive 3D that used to need a specialist could become a prompt.

• Workflows and automation. Huge context plus Google's data and products is where the real leverage would be.

• Price pressure. Cheap, capable models change what is worth building yourself.

My take

We have to wait for the official release. But there is rarely smoke without fire, and if the final model looks anything like these leaks, it will probably surprise us even more.

The lesson for your business is the same either way: don't build on any single model. Build your own data, workflows and stack so you can switch to whichever model wins.

Want an AI setup you own?

I build AI workflows and automation that you own, so you can switch models without starting over. Book a free call and we will look at what makes sense for your business.