A context window is the amount of text — measured in tokens, which are roughly word-sized chunks — that an AI model can consider at one time, and its size decides how much of your conversation, document, or code the model can actually see when it answers.
If your material fits inside the window, the model can use all of it. If it does not, the oldest parts get dropped or the input gets refused, which is why a long chat can suddenly "forget" what you said at the start.
Think of the context window as a desk rather than a filing cabinet. Everything the model reasons about must be laid out on that desk at once: your question, the documents you pasted, the earlier turns of the conversation, and the instructions you gave at the beginning. The model has no memory of what was on the desk yesterday, and no ability to walk over to a cabinet and pull out a file you mentioned an hour ago.
When the desk fills up, something has to come off it. Most tools solve this by quietly discarding the earliest content, which is why a chatbot that followed your formatting rules perfectly at message three may ignore them by message thirty. The mechanism is simple: the window is a hard ceiling on what counts as "in front of" the model.
According to our AI tool database, which holds pricing and capability snapshots for 360 AI tools recorded at verification time, the most recent verification date being 2026-09-18, capability snapshots are one of the attributes captured for each tool — and context limits are exactly the kind of specification that shifts between tools and between versions of the same tool.
Here is a worked example. Suppose you paste a 40-page commercial contract into a chat assistant and ask it to list every clause that mentions automatic renewal. If the whole contract fits in the window, you get a clean list.
If it does not, the tool may either reject the paste with an error, silently cut the document in half, or keep the whole thing but start losing your earlier instructions. The dangerous version is the silent one: you get a confident answer that covers only the first half of the contract, and nothing in the reply tells you that pages 21 through 40 were never read.
The same thing happens with code. Paste a multi-file codebase and ask why a function returns the wrong value, and if the relevant file fell outside the window, the model will still answer — reasoning about the fragments it can see. This is closely related to why AI tools make things up: when the model is missing context, it fills the gap with plausible-sounding invention rather than saying "I can't see that part."
You can read more about that failure mode in our guide to why AI sometimes makes things up and how to tell when it's wrong.
Now the honest trade-offs, because bigger is not automatically better. A larger context window generally costs more per request, since you are paying for the tokens the model processes, and it can slow responses down because there is more material to work through. Size also does not guarantee quality.
A model with a huge window may still pay uneven attention across it — details buried in the middle of a very long input are often used less reliably than details at the start or end. And a large window is not a substitute for good structure: handing a model 200 pages of unedited notes usually produces worse results than handing it 10 pages you have already trimmed to what matters.
The practical decision rule is to match the window to the job rather than chasing the biggest number. For a back-and-forth brainstorm, a modest window is fine. For analysing a long document in one pass, you need a window comfortably larger than the document plus your instructions plus room for the reply.
For anything bigger, split the work into chunks and feed them in sequence, or use a tool built for retrieval, which searches your material and pulls only the relevant pieces onto the desk. One tip that saves real time: before blaming the model for a bad answer on a long input, check whether your content actually fit.
If you pasted more than the tool accepts, the answer you got was about a different document than the one you thought you sent. Pricing and context limits change frequently across tools and versions, so the vendor's own specification page is the only reliable source for a current number — our database snapshot is a starting point, not a live feed.
And remember the limit is shared: your system instructions, the conversation history, and the pasted document all compete for the same space, so a long chat leaves less room for a long document.