Gemini Notebook's new limits count your sources, not your prompts

Google said on Friday that Gemini Notebook will stop counting prompts and start counting compute. From September 2, the research tool's usage limits refresh every five hours instead of once a day, and what each request costs against your allowance depends on four things: how complex the prompt is, how long the chat has run, how many sources the notebook holds, and which features you use.

The announcement, posted by Gemini Notebook product manager Yesul Shin, frames the change as giving users more control over your compute budget. It contains no numbers. Google published no caps, no per-tier multipliers and no unit of account — what it laid out is a mechanism, not a price.

What changes on September 2

• Limits refresh every five hours rather than daily.
• Usage is weighted by prompt complexity, chat length, number of sources and the features used.
• The notebook tracks usage as you go and suggests alternative outputs when the one you asked for would exceed what is left.
• Reaching the limit no longer simply blocks work: heavy outputs such as Video Overviews and Slide Decks can be deferred, and generate automatically later, with optional notifications.
• It applies to consumer accounts on web and mobile.

This already happened to the Gemini app

The mechanism is not new to Google. At I/O in May, the company said it was "moving from daily prompt limits to a 'compute-used' model" in the Gemini app, with limits that "factor in the complexity of your prompt, the features you use and the length of your chat" and refresh every five hours "until you reach your weekly limit." Gemini Notebook inherits the same clock and three of those four factors.

The fourth is new, and it is the one specific to this product: number of sources. Google did not say whether Notebook also sits under a weekly ceiling the way the Gemini app does — the post does not mention one either way.

Why counting sources is the interesting part

Gemini Notebook is source-grounded: answers are assembled from documents you attach, and Google positions it as a research tool rather than a general chatbot. It was renamed from NotebookLM in July, when Google put its usage at more than 30 million people and over 600,000 organizations.

Under a request-count quota, a notebook holding four sources and one holding four hundred cost exactly the same per question. Under a compute-weighted quota they cannot. Every turn against a large corpus does more retrieval and carries more context, and Google is now passing that through to the allowance. The practical effect is that pruning a notebook becomes a budgeting decision and not only an editorial one, and that a long-running chat is more expensive per turn than a fresh one asking the same question.

None of that is new economics. Anyone paying for an API has been metered on context length for years. What is new is that the consumer product now makes the user feel it directly, in a tool whose whole design encourages piling in sources.

More accurate, less legible

Of the two schemes, the compute-weighted one is the more honest. A one-line question and a slide deck synthesised from two hundred documents are genuinely not the same amount of work, and charging them the same was always a rough approximation. The problem is what replaces it. Token billing on an API is legible: the unit is public, countable before you send anything, and priced on a published rate card. Google's compute unit is none of those. There is no stated unit, no conversion and no way to know in advance what a given prompt will draw down.

In place of a rate card, Google is offering instrumentation — a running usage view and a system that proposes a cheaper output when the expensive one will not fit. For someone who just wants to keep working, that is a reasonable substitute. For anyone trying to plan a workload against a known budget, it is not. And with no before-and-after figures published, there is no way to say yet whether a given user ends up with more usable capacity than under the daily caps or less. That only becomes visible once the rollout starts.

The deferral feature deserves its own note. Converting "you have hit your limit" into "we will run it later and tell you when it is ready" is a load-shifting tool for Google at least as much as a convenience for the user: it moves expensive multimodal generation off the peak and into whatever capacity is free. As consumer AI products keep running into the cost of the compute underneath them, that shape — queue it rather than refuse it — is likely to become more common than the hard cap it replaces.