What is a context window?
How much text an AI model can read at once, how tokens relate to words, why replies have their own limit, and what huge context windows can and cannot do.
Last updated Oct 1, 2026
Every time you send a message to an AI model, it reads one block of text: your message, the conversation so far, any files you attached, and instructions written by the app. The context window is the maximum size of that block. Think of it as the model's working memory for a single request. Anything inside it can be used. Anything outside it does not exist as far as the model is concerned.
Context windows are measured in tokens, not words or pages, so it helps to start there.
Tokens, not words
Models do not read letters or whole words. They read tokens, which are chunks of text from a fixed vocabulary built before the model is trained. A common short word like "the" is usually a single token. A longer or rarer word may be split into pieces, so "unbelievable" might become something like "un", "believ" and "able". Spaces, punctuation and digits use tokens too.
A handy rule for English prose: one token is about 0.75 words. Put the other way, 100 words is roughly 133 tokens. That makes quick conversions easy:
- 8,000 tokens is about 6,000 words, a long essay.
- 128,000 tokens is about 96,000 words, roughly a full novel.
- 1,000,000 tokens is about 750,000 words, a shelf of novels or a sizeable codebase.
The rule is only an estimate. Each model family uses its own tokenizer, so the same paragraph can produce different counts on different models. Code, tables and many languages other than English usually need more tokens per word. When the exact number matters, use the provider's token counting tool.
Input limits and output limits
Two numbers usually appear on a model's spec sheet. The context window is the total space for a request. The maximum output is the most the model can write in a single reply. On most models the two share one budget: your prompt plus the reply must fit inside the window, and the reply on its own must also stay under the output cap.
The output cap is often much smaller than the window. Claude Opus 5.5 can read up to 1,000,000 tokens but writes at most 128,000 in one reply. Llama 3.3 70B pairs a 128,000 token window with a 4,096 token output limit, so it can read a long report but would need many turns to rewrite it. A few models list matching numbers. Kimi K3 shows 1,048,576 for both, which means a reply could in principle fill the whole window, although whatever you send still takes up part of that same space.
Reasoning models add a wrinkle. When a model works through a problem before answering, those hidden reasoning tokens usually count toward the output limit, even though you never see them. A hard question can use up a large share of the output budget before the visible answer starts.
How context windows compare
Not long ago, a few thousand tokens was normal. Today, around a million is common at the top end. Here is how a few models in the Modeldle catalog compare:
| Model | Context window | Max output | Window in words, roughly |
|---|---|---|---|
| GPT-4 | 8,192 | 8,192 | 6,000 |
| GPT-4o | 128,000 | 16,384 | 96,000 |
| Claude Opus 4.5 | 200,000 | 64,000 | 150,000 |
| Grok 4.7 | 500,000 | 500,000 | 375,000 |
| Gemini 3.1 Pro | 1,048,576 | 65,536 | 786,000 |
| GPT-6 Sol | 1,050,000 | 128,000 | 788,000 |
If 1,048,576 looks oddly precise, it is 1,024 times 1,024, a "binary million" that engineers like because it lines up with powers of two. For practical purposes it is a one million token window.
Why a big window is useful
- Whole documents at once. You can paste a full contract, a research paper or a year of meeting notes and ask about all of it, instead of feeding it in pieces and hoping nothing important falls between them.
- Large codebases. Coding assistants can see many files together, which helps them follow how a change in one place affects another.
- Long conversations. Chat apps keep earlier messages in the window. When a chat outgrows it, the app has to drop or summarize older turns, which is why a very long chat can "forget" something you said at the start.
The limits of a big window
Cost and speed
You pay for every input token, on every request. Sending a full 1,000,000 token prompt to Claude Opus 5.5 costs $4 at its listed price of $4 per million input tokens, and you pay it again each time you send it. The same prompt on Gemini 3.8 Flash, listed at $0.75 per million, costs 75 cents. Long prompts also take longer to process, so the first words of the reply arrive later. The guide to token pricing walks through the arithmetic.
Lost in the middle
Fitting text into the window does not guarantee the model uses it well. A 2023 study titled Lost in the Middle found that models answered best when the relevant fact sat near the start or end of a long prompt, and noticeably worse when it was buried in the middle. Newer models handle long inputs far better, and labs now test this with "needle in a haystack" style benchmarks. Still, recall across a huge prompt is less reliable than across a short one, and connecting many scattered facts is harder than finding a single one.
A few habits help: put the most important instructions at the start or the end, leave out material the model does not need, and ask it to quote the passage it relied on so you can check.
A window is not memory
The context window starts fresh with every request, and nothing in it permanently changes the model. Apps that seem to remember you are saving notes and quietly adding them back into the window. It is also separate from what the model learned in training, which stops at its knowledge cutoff.
In the quiz
Modeldle often puts two model cards side by side and asks "Which can read more text at once?", or asks how big one model's context window is. Both are context window questions. A sister question, "Which can write a longer reply?", is about the max output instead, so read carefully. Remember that the largest windows in the catalog sit around one million tokens, that older models like GPT-4 were measured in thousands, and that the output cap is a separate, usually smaller number. Ready to try? Play today's quiz.