Context Window Explained: How Much an AI Holds at Once

A context window is the maximum amount of text a model can have in front of it during a single request, measured in tokens. Everything is in there: your question, the document you pasted, the system instructions, every previous turn of the conversation, and the model’s own reasoning. When it’s full, something has to go.

Grok 4.5, released July 8, 2026, shipped with a 500,000-token window. Claude’s current models carry 1 million tokens. A million tokens is roughly 750,000 words, or about eight copies of Moby-Dick. Which sounds like the end of the problem.

It isn’t, and the reason is the most useful thing in this article: a bigger window doesn’t mean the model pays equal attention to all of it.

Key Takeaways

• The context window holds your prompt, pasted documents, conversation history, system instructions, and the model’s reasoning, all counted in tokens together.
• Grok 4.5 (July 8, 2026) has a 500K-token window with native video input. Claude 4.6 and later models carry a 1M-token window at standard pricing, with no premium for long requests.
• Liu et al. (TACL, 2023) found a U-shaped performance curve: models do best when the relevant information sits at the beginning or end of the context and “significantly degrades” when it’s in the middle.
• That U-shape holds even in models explicitly built for long contexts, which is why “just paste everything” is worse advice than it sounds.
• Practical rule for kids: put the important thing first or last, and never assume the model read page 40 of 80 as carefully as page 1.

What’s actually in the window

The window is not a memory. It’s a workspace, rebuilt from scratch on every request. Nothing carries over unless it’s resent.

That means on turn 15 of a conversation, the model isn’t remembering turns 1 through 14. Those turns are being sent again, as text, filling up the window. This is why long chats eventually slow down, get expensive, or start losing details from the beginning. And it’s why prompt caching exists as a way to make that resending cheap.

Five things compete for the same space:

1. The system prompt. Instructions from whoever built the tool. Often thousands of tokens, invisible to the user.
2. Tool definitions. If the model can search the web or run code, the descriptions of those tools sit in the window too.
3. Pasted content. Documents, code, images, transcripts.
4. Conversation history. Every previous turn, both sides.
5. The model’s reasoning. On models with extended thinking, deliberation occupies window space as it’s generated.

Anthropic’s documentation notes that for Claude 4.6 and later, the full 1M token context window comes at standard pricing: “A 900k-token request is billed at the same per-token rate as a 9k-token request.” That’s a meaningful change from earlier long-context pricing, which often carried a premium.

Grok 4.5’s specs, released July 8, 2026, list a 500K context window alongside reasoning, multimodal input, real-time knowledge from X, and native video input. Grok 4.6 followed on August 12, 2026, per the running chronology of 2026 AI releases. For comparison, the September 1, 2026 Fable 5.1 release reported a 1M-token window with 128K maximum output.

Why the middle gets lost

Here is the research finding that should change how your family uses long documents with AI.

Nelson Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang published “Lost in the Middle: How Language Models Use Long Contexts” in Transactions of the ACL. Their core result, in their own words: “performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts.”

They tested this two ways: multi-document question answering, where the answer sits in one of many documents, and key-value retrieval, where the model has to find a specific pairing. In both, accuracy plotted against position produced a U shape. High at the front. High at the back. A valley in the middle.

And the crucial detail: this held even for models explicitly designed for extended context. Having room for a document is not the same as using all of it well.

Why does this happen? The honest answer is that it isn’t fully settled. The attention mechanism at the heart of every one of these models comes from Vaswani et al.’s 2017 paper, and how attention distributes across very long sequences is still an active research question. Contributing factors likely include how attention is distributed across positions during training, the fact that training data has important information disproportionately at the start and end of documents, and positional encoding behavior at long ranges. Anyone who gives you a confident single-cause explanation is overstating what’s known.

What matters practically is the shape of the curve, not its cause. If your kid pastes an 80-page PDF and asks about something on page 40, the answer is less reliable than if the same fact were on page 2 or page 79. That’s a real, measured effect, not a hunch.

Window size vs. usable attention

The industry has spent three years racing on window size. It’s a clean number to advertise, and bigger is genuinely better for some things: analyzing an entire codebase, reading a full novel, holding a long agent session. Our comparison of Grok 4.5 and 4.6 for teens covers what the 500K window actually enables.

But window size and usable attention are different properties, and only one gets advertised. A model with a 1M-token window and a strong U-shaped attention curve will do worse on a 500-page document than a model with a 200K window that you’ve chunked intelligently.

The practical techniques that follow from this are simple enough for a middle schooler:

Chunk deliberately. Instead of pasting 80 pages and asking one question, paste the relevant 10 pages. Less context, better attention, cheaper.

Put the question at the end. If the model does best at the ends, and your question is the thing you most want attended to, put it last. This is genuinely why many prompt guides tell you to repeat your question after a long document.

Repeat critical constraints. If there’s a rule that must not be violated (“cite only from the provided text”), state it at the top and again at the bottom.

Start a new conversation on a new topic. A fresh window is cleaner and cheaper than one carrying 40 turns of unrelated history.

School tools are starting to do the chunking for you. Google’s August 2026 Classroom update pulls in the title, assignment instructions, and curriculum materials for one specific assignment rather than an entire course, which is a deliberate choice about what belongs in the window.

How to Teach Your Kid About Context Windows

The window is a desk, a whiteboard, or a backpack. Whichever object your kid actually uses, use that.

Ages 5–8: The lunch tray

Give your kid a lunch tray or a small box and a pile of toys. Ask them to fit everything on the tray. They can’t. Ask which things they’d leave behind. Then say: “Computers that answer questions have a tray like this. They can only look at so much stuff at one time. If you give them too much, some falls off.” Then the second half: put ten toys on the tray, ask them to close their eyes, and ask which ones they remember. They’ll name the first and last ones. Point that out. That’s the U-shape, discovered by a six-year-old.

Ages 9–12: Hide the fact

Write twenty short facts on a numbered list. Read the list aloud to your kid once, at normal speed. Then quiz them on fact 3, fact 11, and fact 19. Almost every kid nails 3 and 19 and misses 11. Now tell them that Liu et al. found exactly the same pattern in language models, published in a 2023 paper, and that it’s called “lost in the middle.” Then ask the useful question: if you had to hide an important fact where the model would find it, where would you put it?

Ages 13+: Run the experiment

Have your teen build a real test. Take a long article, roughly 3,000 words. Insert a made-up but specific sentence (“The committee met on a Tuesday in Cartagena”) at three positions: near the start, the exact middle, and near the end. Run three separate requests, each with one placement, asking the model to find that sentence. Record hit or miss. Repeat with a longer article. They’re replicating the lost-in-the-middle finding with a sample size of six, which is small but real, and they’ll understand the paper afterward in a way no summary delivers.

The question to ask: “If you can only look at one page at a time, how would you find one sentence in a hundred pages?”

Model to window: what fits and what it means

Model	Context window	Roughly equivalent to	What it’s actually good for
Claude 4.6 and later (incl. Fable 5.1, Opus 5)	1,000,000 tokens	~750,000 words, ~8 novels	Whole codebases, long agent sessions; standard per-token pricing across the full window
Grok 4.5 (July 8, 2026)	500,000 tokens	~375,000 words	Long documents plus real-time X data and native video input
Claude Sonnet 4.6 and earlier	Smaller windows, older tokenizer	Varies	Most homework-scale tasks; cheaper per token
A 20-page school PDF	~15,000 tokens	15 pages of dense text	Fits in essentially any modern model with room to spare
A 300-page novel	~124,000 tokens	1 novel	Fits in a 500K or 1M window, but expect middle-of-book questions to be less reliable
40 turns of chat history	20,000–100,000 tokens	Varies wildly	The invisible thing filling your window; start fresh for new topics

Token estimates use the documented rule of roughly 4 characters per token. Note that Anthropic’s docs flag that Claude 4.7 and later use a newer tokenizer producing about 30% more tokens for the same text, so a “1M window” holds somewhat less prose than the arithmetic suggests. Our tokens explainer covers that wrinkle.

What to do at home

Teach “paste less, ask better”

The instinct is to paste everything. The better move is to paste the relevant section and put the question after it. This costs less, produces better answers, and requires your kid to actually locate the relevant part, which is the schoolwork.

Start a new chat when the topic changes

Old turns keep occupying the window and can bleed into new answers. “New topic, new chat” is a one-second habit that prevents a category of weird responses.

Check middle-of-document answers twice

If your kid asks about something buried in a long source, verify it against the source. This is the single highest-yield verification habit, because it targets the documented failure mode rather than a vague worry about hallucinations. Our guide on when AI is wrong covers the broader category.

Don’t confuse the window with memory

A context window resets every request. If a chatbot appears to remember something from last week, that’s a separate stored-memory feature, not the context window, and it’s worth knowing which one your kid’s tool has.

What not to do

Don’t treat a bigger window as a straightforward upgrade. The U-shaped curve from Liu et al. held even in models built for long contexts, so a 1M-token window is permission to paste more, not a promise that all of it will be used well. The skill of choosing what to include did not go away.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can explain that the window holds the question, the document, and the whole chat history together, and that the middle is the weak spot.
• Month 2 red flags: They paste entire textbooks and trust the answers. Or they keep one endless chat running across every subject.
• Month 3 self-check: Ask them where to put the most important instruction in a long prompt. “First or last” with a reason is a pass.

Frequently Asked Questions

Does a bigger context window mean the AI is smarter?

No. It means the AI can look at more text at once. Liu et al. (2023) showed that models use information at the beginning and end of a long context much more reliably than information in the middle, and that pattern held even in models designed for long contexts.

How much is a million tokens in pages?

Roughly 750,000 words, or about 2,500 to 3,000 pages of typical prose. Tokenizers vary, though: Anthropic’s docs note that Claude 4.7 and later produce about 30% more tokens for the same text than earlier models, so the practical page count is lower than the raw arithmetic.

Why does my kid’s chatbot forget things from earlier in a long conversation?

Either the earlier turns fell out of the window as newer content pushed in, or they’re still present but sitting in the low-attention middle. Both produce the same experience. Starting a new conversation for a new topic avoids it.

Is a 500K window enough for schoolwork?

Comfortably. A 20-page PDF is around 15,000 tokens, so essentially any current model has room. Window size is rarely the constraint for homework; attention placement and cost are the real variables.

Should we paste the whole textbook chapter or just the relevant pages?

Just the relevant pages, for three reasons: better attention on a shorter context, lower cost, and the fact that locating the relevant section is itself the learning. Pasting everything outsources the part of the work that builds skill.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2023). “Lost in the Middle: How Language Models Use Long Contexts.” Transactions of the Association for Computational Linguistics. https://arxiv.org/abs/2307.03172
2. Anthropic. (2026). “Pricing.” Claude Platform Documentation. https://platform.claude.com/docs/en/about-claude/pricing
3. Anthropic. (2026). “Prompt caching.” Claude Platform Documentation. https://platform.claude.com/docs/en/build-with-claude/prompt-caching
4. MarkTechPost. (2026, September 1). “Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1.” https://www.marktechpost.com/2026/09/01/anthropic-releases-claude-fable-5-1-and-claude-mythos-5-1-52-6-on-terminal-bench-science-and-75-cheaper-cache-reads/
5. Wikipedia. “2026 in artificial intelligence.” (Grok 4.5 release, July 8, 2026; Grok 4.6, August 12, 2026.) https://en.wikipedia.org/wiki/2026_in_artificial_intelligence
6. Google. (2026, August). “Google AI updates, August 2026.” The Keyword. https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/
7. Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). “Attention Is All You Need.” NeurIPS 2017. https://arxiv.org/abs/1706.03762