Context Window Calculator

What a context window is

A context window is the maximum amount of text an AI model can consider at one time, measured in tokens. It is best understood as the model’s short-term working memory. On every single turn, that window has to hold four things at once: your prompt, any files or documents you paste in, the conversation so far, and the reply the model is about to write. When people say a model “has a 200K context window,” they mean all of that together cannot exceed roughly 200,000 tokens.

Unlike a human’s memory, this window is hard-edged. The model does not gradually forget older details — everything inside the window is available with equal access, and everything outside it simply does not exist as far as the model is concerned. That is why context size is one of the most practical numbers to know about any model: it decides, precisely, how much you can show it before something has to be left out.

Tokens vs words vs pages

Context windows are quoted in tokens, but most of us think in words or pages. The calculator above converts between them using the same rules of thumb our token counter uses for English prose:

• 1 token ≈ 4 characters of English text
• 1 token ≈ ¾ of a word — so 1,000 words ≈ 1,300 tokens
• 1 page (~500 words) ≈ 650 tokens

Chain those together and the headline numbers become concrete. A 128K window is around 200 pages; 200K is around 300 pages, roughly a full-length novel; and 1M tokens is around 1,500 pages, or several books at once. Bear in mind these are averages for ordinary prose. Source code, JSON, spreadsheets, URLs, emoji and languages other than English pack more tokens per character, so they fit fewer pages into the same window. To count a specific document precisely, paste it into the token counter first.

Why a bigger window isn’t always better

It is tempting to treat context size as a simple “bigger is better” spec, like storage on a phone. In practice there are three catches.

First, attention is uneven. Research on long-context models keeps finding the same pattern, often called “lost in the middle”: models recall information placed at the very start or very end of a long input far more reliably than material buried in the middle. Filling a million-token window to the brim does not guarantee the model actually uses every part of it well.

Second, cost scales with tokens. Almost every API bills per token, so a prompt that stuffs 500K tokens of context into each call is dramatically more expensive than one that sends a tight 5K. Third, latency grows too: the more tokens a model has to read before it starts answering, the longer you wait for the first word. A focused, well-chosen context is usually faster, cheaper and more accurate than an enormous unfocused one.

Long context vs RAG

Once your material is bigger than a comfortable fraction of the window, you face a choice: paste everything and let the long context handle it, or retrieve only the relevant pieces and feed those in. The second approach is called retrieval-augmented generation, or RAG.

Long context wins when the task genuinely needs the whole document at once — summarising a contract end to end, reasoning across a full report, or holding a long conversation in mind. It is simple: no extra infrastructure, just paste and ask. RAG wins when you have far more material than any window could hold (a knowledge base, years of documents, an entire wiki) and each question only touches a small slice of it. Instead of sending everything every time, a retrieval step finds the handful of relevant passages and sends just those — cheaper, faster, and sidestepping the “lost in the middle” problem. Many real systems blend both: retrieve the right chunks, then rely on a large window to reason over them comfortably.

How much really fits

Consider a worked example. Say you want a model to read a 300-page book. At roughly 500 words a page, that is about 150,000 words, which at ¾ of a word per token comes to around 195,000 tokens. What can hold it?

• A 128Kmodel (GPT-4o, a typical Llama build) cannot — the book is about 1.5× too large, so you would need to split it or switch to RAG.
• Claude at 200K just fits it, with only a sliver of room left over — not enough headroom to also paste a second book or a long question, so plan for the reply.
• Gemini at 1M swallows the book with room for four or five more like it, which is why million-token windows are pitched at whole-codebase and multi-document work.

The lesson from the calculator is always to leave headroom. If a 200,000-token document exactly fills a 200,000-token window, there is no space left for the model to write its answer. A safe habit is to keep input under about 70–80% of the window, so 30–40K tokens stay free for the response. To weigh windows against price, speed and capability across providers, compare them in our AI model comparison.