What is a context window?
Plain-English definition · Updated 2026-10-06. Numbers dated; verify with the vendor.
AITop is an independent guide. Prices and features change fast — always check the vendor's page.
The context window is how much a model can hold in mind at once, measured in tokens: your prompt, every document you attach, the whole chat history, and the reply being written all share one space. It is the model's workbench, not its brain — nothing outside the window exists for the current answer. Advertised windows now reach into the hundreds of thousands and even millions of tokens, but the usable window — where quality stays sharp — is smaller, and that gap is where marketing lives.
What it actually means
Think of it as one shared whiteboard of fixed size. Your instructions go on it, the PDF you uploaded goes on it, the last forty messages of the conversation go on it — and when it fills, something has to come off. Older turns get compressed, summarized or silently dropped, which is why a chatbot that "remembered" something yesterday can forget it today: the memory left the whiteboard, not the model.
Two consequences follow. First, the window is shared: a 200k-token window with a 150k-token document attached leaves less room than you think, and long conversations quietly eat it. Second, quality is not uniform across the board — models attend most reliably to the beginning and end of what they can see, and long-buried details are where mistakes breed.
Why it matters when you're picking a tool
Because the workloads that feel "big" — a 300-page contract, a whole codebase, a semester of chat history — live or die on this number and on how the tool manages it. The right question is never "how big is the window?" but "what does the tool do when the window runs out?" Vendors diverge hard here: some compress and summarize automatically, some retrieve only relevant chunks (see RAG), some simply truncate and hope you don't notice. Same advertised number, very different experience.
The 2026 reality check
Window sizes became a marketing arms race after million-token claims landed, so the number stopped being informative on its own. What distinguishes tools now is effective use: how much of a huge document the model can actually reason over without dropping the needle you asked about. Benchmarks that test long-document comprehension show large gaps between models with identical claimed windows — and long-context input is also billed by the token, so a careless 500k-token prompt is a real invoice. The practical hierarchy in 2026: a modest window with excellent retrieval beats a giant window with lazy truncation, almost every time.
Quick checklist
- Test with your document, not the vendor's demo — upload your longest real file and ask about a detail buried in the middle.
- Ask what happens at overflow: compression, retrieval, or silent truncation?
- Check the price of long prompts if you pay per token — window size and bill size are the same knob.
- For recurring large-document work, prefer tools with explicit retrieval over tools that just claim a big window.
Where you'll hit it
Window claims and their fine print run through the best AI chatbots, ChatGPT vs Claude (two different philosophies of long-context handling), and Notion AI vs ChatGPT (where workspace search does the heavy lifting). Related terms: RAG and AI token — or back to the full glossary.
The bottom line
The context window is capacity, not competence. Treat advertised token counts as the size of the parking lot, not the quality of the driving — and judge tools by what they do when the whiteboard fills. That one question separates the tools engineered for real work from the ones engineered for spec sheets.