What fits in a context window?
Model makers publish the window for each model in their developer documentation, and in 2026 the common sizes run from about 128,000 tokens to a million or more. Using OpenAI's rule of thumb of 100 tokens for every 75 words of English, this is what a 128,000-token window holds.
| Document | Words | Tokens, roughly | Fits in 128,000? |
|---|---|---|---|
| A long email | 400 | 530 | Easily |
| A 20-page commercial lease | 9,000 | 12,000 | Easily |
| An 80-page strata report | 32,000 | 43,000 | Yes, with room for questions |
| A full-length novel | 90,000 | 120,000 | Only just |
| A year of a busy inbox | Millions | Millions | No |
The window is shared. The app's own hidden instructions, the document you pasted and a long reply all draw on the same allowance.
What happens when a chat outgrows the window?
Something gets left out. A model keeps nothing between messages: each time you send one, the app sends the whole conversation again, and when that no longer fits, the app has to choose what to drop. Most drop the oldest messages or swap them for a short summary.
That is why a long chat can lose an instruction you gave at the start. The instruction is simply no longer in what the model is reading. Starting a fresh chat with the important facts at the top fixes it more reliably than repeating yourself 40 messages in.
Does a bigger window mean every page is read equally well?
No, and the research on this is specific. In a 2023 study called "Lost in the Middle", Nelson Liu and colleagues at Stanford found that language models used information at the start and the end of a long input much better than information buried in the middle. Models have improved since, but the habit is still worth working around: put the instruction that counts first or last, and when you need one detail from a long document, name the section you mean.
What does it cost to question an 80-page strata report? (example figures)
A real estate agency in Port Macquarie receives an 80-page strata report for a unit it is listing, about 32,000 words or 43,000 tokens. The report fits comfortably in a 128,000-token window, so the agent attaches it and asks 10 questions: the sinking fund balance, any building defects, the pet by-law.
Each question re-sends the whole report, so 10 questions come to 430,000 input tokens. At example rates of $4.50 per million input tokens and $22.50 per million output, with 300-token answers, the session costs about $2. The money is not the risk. The special levy for concrete cancer repairs is on page 41, deep in the middle, and the answer to "Are there any special levies?" misses it. Asking "What does the levies section say about special levies?" finds it. For anything that goes into a listing or a buyer's hands, the agent still checks the page.
Does Cora need you to paste in your reports?
No, CoraVision does not depend on you pasting reports into a chat: it reads Google Ads, Meta and Google Analytics 4 itself, first when you connect them and then every night. The direction lives in the plan, which outlasts any one conversation. Cora builds a 12-month strategy around the one number you choose to chase, each strategy becomes a project with an action plan, and the weekly check-ins pick up from there. Transcripts of your calls and chats are kept with your account.
Questions
Is a context window the same as an AI's memory?
Not quite. The window is what the model can see in a single request. Memory features in apps save notes about you and insert them into the window at the start of a new chat, so the memory belongs to the app, and the model itself keeps nothing between requests.
Why did the AI ignore something I told it earlier?
In a long conversation the earlier message may have been dropped or summarised to make room, or it may sit in the middle of a long input, where models pay least attention. Repeat it near the end, or start a fresh chat with it at the top.
Does using more of the window cost more?
Yes, when you pay by the token. Every request is charged on the tokens sent, so a question about a 40,000-token document costs far more than the same question on its own, and the whole document is charged again with each follow-up.