The rumor mill has already moved on to Gemini 4. The headline number floating around: a 10-million-token context window. Before anyone starts pasting entire codebases into the prompt box, it's worth separating what Google has actually confirmed from leak-season wishful thinking, and Geeky Gadgets has a good breakdown of both.

On the confirmed side, Google has said Gemini 4 is in pre-training. That's a real, verifiable statement, and it matters. Pre-training is the long, expensive phase where the model ingests a huge chunk of its training data before fine-tuning and alignment work even starts. It takes months. So a model in pre-training right now is a model you shouldn't expect to touch for a while, no matter what optimistic timelines get thrown around.

The 10-million-token context window, on the other hand, is unverified. No benchmark leaks, no API documentation, nothing you can actually check. It's the kind of number that sounds plausible precisely because context lengths have been climbing steadily, so you extrapolate the trend line and out pops a big round figure that leaks love. I'd treat it the way I treat a "works on my machine" bug report: interesting, unfalsifiable, and probably wrong in the details even if the direction is right.

Why does anyone even want 10 million tokens? As a developer I can tell you the appeal is real. Current context windows already let you load a decent chunk of a codebase into a model, but you still hit walls with large monorepos, long transcripts, multi-document analysis. Whatever the exact word count shakes out to, we're talking entire book series fitting in a single prompt, with room left over for your follow-up questions. The workflow shift would be genuine: stop cherry-picking what the model sees, and just hand it everything.

But there's a catch, and it's the same catch that's existed since context windows started ballooning. Effective context is not the same as advertised context. A model that technically accepts 10 million tokens but only reasons well over a fraction of that is a spec sheet, not a feature. Attention quality degrades, retrieval gets fuzzy in the middle of the window, and latency plus cost scale with every token you feed it. Google would need real architectural improvements, probably some form of sparse or hierarchical attention, not just a bigger buffer.

So where does that leave us? Pre-training is confirmed, timing is early, and the 10-million-token figure stays a rumor until an API changelog says otherwise. I'll be watching for the first sign of real evidence, anything checkable like documentation or benchmarks, before I take the number seriously. Models, like deployments, eventually reveal themselves through their receipts.