Why One Million AI Tokens Still Don’t Replace a Good Note-Taking System
One million tokens sounds enormous. For anyone building a “second brain” in tools like Obsidian, Notion, Joplin, or similar apps, it raises an obvious question: if modern AI models can read huge amounts of text, why bother organizing notes at all?
Why tag them? Why link them? Why build folders, indexes, or search systems when you could simply throw your entire knowledge base into an AI prompt and ask questions?
The short answer: because one million tokens is less room than many people think, especially outside English. More importantly, every question can require the model to read the same mountain of text again. That can become expensive, slow, and less accurate than working with a smaller set of relevant notes.
A million tokens is not always 750,000 words
The common rule of thumb says one million tokens equals about 750,000 words. That estimate is often repeated, but it is not universal.
Tokens are not the same as words. They are small chunks of text created by a tokenizer, the system that breaks language into pieces the AI can process and bill for. A token might be a whole word, part of a word, punctuation, or even a space-like structure depending on the model and language.
That means the number of words you can fit into one million tokens depends heavily on the language and tokenizer.
A comparison using the same article in German and English shows how big the difference can be. The German version contained 1,134 words, while the English version contained 1,276 words. Both texts had the same content and included only the main article text and subheadings.
Using modern OpenAI tokenization, one million tokens works out to roughly 543,000 to 649,000 German words, depending on the tokenizer used. That is far below the popular 750,000-word estimate.
For English, the estimate is closer, and in some cases even conservative. For German, however, it can be far too generous.
The tokenizer matters more than you might expect
The same German text produced very different token counts depending on the tokenizer. With one encoding, it became 1,747 tokens. With another, it became 2,088 tokens.
That is a difference of about 20 percent from the same text, simply because the splitting method changed.
For English, the difference was almost nonexistent: about 1.200 versus 1.202 tokens per word. This suggests that newer tokenizers have improved efficiency mainly for non-English languages, while English was already handled efficiently.
For AI tools from companies that do not publish their tokenizers, exact comparisons are harder. Some providers also change tokenization between model generations, which can increase the number of tokens required for the same text. So any cost estimate should be treated as a realistic approximation rather than a perfect calculation.
What happens with a real note collection?
A personal knowledge base can grow faster than expected. A collection of around 470 notes may easily reach 350,000 to 400,000 words over several years.
In theory, that still fits inside a large AI context window. But only just. And the collection keeps growing.
The bigger issue is that a context window is not permanent memory. When you ask a question through an API, the model does not magically remember everything from last time. The relevant text has to be sent again, processed again, and paid for again.
For example, with GPT-5.6 Sol, one million input tokens cost five US dollars. Once a single request passes 272,000 tokens, the price rises to ten dollars per million input tokens.
A German note collection of about 400,000 words may equal roughly 616,000 tokens. Because that sits above the higher pricing threshold, one full-collection query would cost around $6.16.
That is not for a month. That is for one question.
Ask ten questions in an afternoon, and the cost can reach around $60. By comparison, even a premium productivity app subscription may cost far less per month.
There are two important caveats. First, this applies to API usage, where developers pay directly per token. Consumer subscriptions such as ChatGPT Plus or Claude Pro usually do not charge per token in the same way, though they do have usage limits. Second, caching can reduce the cost. If the same large input is reused and qualifies for a cache hit, some providers charge only a fraction of the normal input price.
In the example above, caching could reduce $6.16 to about 62 cents. But cache windows are often short, sometimes only minutes to an hour, depending on the setup. That makes caching useful, but not a complete solution.
Why organized notes still matter
The point of a note-taking system is not to compete with AI. It is to make AI more useful.
A good system reduces what the model needs to read. Instead of sending 400 notes into the prompt, you send the ten that actually matter.
That can happen through simple full-text search, tags, links, folders, or a semantic search index that finds related material by meaning rather than exact keywords.
The difference is dramatic.
If the whole collection is about 616,000 tokens, a set of ten relevant notes might be around 9,000 tokens. That changes the cost from around $6.16 to about five cents.
In token volume, that is roughly a 67-to-1 reduction. In price, the gap can be even larger because the smaller request stays below the higher-cost threshold.
There is also a quality benefit. An AI model that reads ten focused notes is often more accurate than one forced to scan hundreds of loosely related notes. Large context windows are impressive, but more context is not always better context. Too much irrelevant material can dilute the answer.
The real limit may not be the AI context window
For everyday users, the limiting factor is often not the theoretical token limit at all. It is the product interface.
Consumer AI tools usually restrict how many files, projects, notebooks, or sources you can upload. The exact numbers vary by plan and can change over time.
For example, some ChatGPT project limits depend on whether you use a free, Plus, Pro, Business, or education plan. Gemini Notebook also sets source limits depending on the plan, while each source can be quite large.
This changes the real question. It is not only “Can the model understand one million tokens?” It is also “Will the tool let me upload my knowledge base in the first place?”
That is where your own note system becomes valuable. It does not need to upload everything every time. It helps you decide what should be shown to the AI for a specific question.
The downsides of maintaining a second brain
A second brain is not free. It costs time, discipline, and maintenance. Tools also have practical limitations.
Obsidian and Joplin work on phones, but background syncing can be limited depending on how you set them up. If the mobile app is closed before syncing finishes, a note written on the go may not appear on your computer until you open the app again. On desktop this is less noticeable because the app is usually running.
Obsidian’s own sync service requires the app to be active for syncing. If you store the vault in a cloud folder or use another sync service, that external system may handle file transfer instead.
Notion has its own limits, including file-size restrictions on some plans. Gemini Notebook has another concern: deleting a notebook can remove its sources as well, while conversations may be handled separately.
These details matter because a note system is only useful if it fits the way you actually work.
So, do you still need Obsidian, Notion, or another note app?
If you only have a few dozen notes, probably not. A simple folder, document collection, or basic search may be enough.
But if you have hundreds of notes, research clips, meeting summaries, drafts, quotes, ideas, and project documents, organization quickly becomes valuable. Not because AI cannot read large amounts of text, but because it should not have to read everything for every question.
The smartest workflow is not “AI replaces notes.” It is “notes feed AI the right context.”
A well-organized knowledge base helps you find the right material faster, reduces AI costs, avoids upload limits, and improves answer quality. Large context windows are powerful, but they are not a replacement for structure.
The practical advice is simple: collect notes for two weeks and see what happens. If you end up with 30 notes, keep it simple. If you end up with 300, you will quickly understand why search, tags, links, and structure still matter.
One million tokens may sound like unlimited memory. In practice, it is a large but costly reading window. A good note system turns that window into something far more useful: focused context.
