Let’s address something that feels dangerously close to false advertising.
Official documentation and public messaging repeatedly highlight that Gemini Pro supports a 1 million token context window. That number is heavily emphasized. It is one of the key selling points.
But in the Gemini app, the real-world experience does not match that claim.
After long roleplay and creative sessions - and I have done this extensively over months - the model begins to forget major events far, far earlier than 1M tokens. Not slightly earlier. Dramatically earlier.
In my testing, the model begins losing memory coherence somewhere around 150k–200k tokens. That is not a small deviation. That is roughly one sixth of the advertised capacity.
At around 120k tokens back in the conversation, the model had already forgotten critical events entirely. Not minor details. Entire story arcs.
Meanwhile, in Google AI Studio, I can clearly see the token usage. I can reach 500k+ tokens, and the model still remembers what happened at the very beginning. It can even quote specific phrases from early messages.
The difference is not subtle.
So what is happening in the Gemini app?
If the system is aggressively “optimizing” context by truncating, summarizing, or silently compressing earlier messages, then the practical context window is not 1M tokens in any meaningful sense.
It becomes a theoretical number, not a functional one.
For creative users who rely on long-form continuity - writers, roleplayers, researchers - this is not a minor inconvenience. It fundamentally breaks the experience.
When a product is marketed around a 1M token context window, but in practice behaves as if it has a fraction of that, users are going to question the claim.
Is the 1M token window fully available in the Gemini app, or is heavy internal optimization reducing effective memory?
If optimization is happening, why is this not clearly communicated?
Right now, it feels like the advertised capability and the actual behavior in the Gemini app are not aligned.
And that gap deserves a clear explanation.