What is a context window and why does it change what a model can do?

Model announcements lean hard on this number: a million tokens of context, printed like a top speed. It is a real and useful figure, but it measures something narrower than the marketing suggests. This guide explains what the window actually holds, shows the sizes the vendors publish, read from their own pages and kept in our ledger, and covers the part the announcements skip: what filling a window costs, and what happens at the edge of it.

Tokens, the unit the window is measured in

A token is the chunk of text a model processes at a time: a short word, a piece of a longer one, a punctuation mark. English averages out near three quarters of a word per token, so 200,000 tokens is roughly the length of one long novel. A context window is the maximum count of tokens the model can consider in a single request: everything it reads and everything it writes back, together, inside one limit. It is working memory, not knowledge. The model’s training does not sit in the window; your conversation does.

What fills the window

The window holds more than your typing. Each request packs in the system prompt, the standing instructions the vendor or app developer wrote; the conversation so far, resent in full with every turn; anything you attached or pasted; whatever a search or file tool returned this turn; and the space the reply itself will occupy. A short question at the end of a long chat is not a short request, because the chat rides along with it.

That is why long conversations degrade. When the total no longer fits, the software in front of the model quietly drops or summarises the oldest turns. The model answers confidently without them, and the person on the other end experiences it as forgetting.

The published windows, from our ledger

These figures come from the vendors’ own model and pricing pages, each read on 6 September 2026 and recorded in our pricing ledger with its source. Not all of them sit on a price page. Google prints the Gemini window on the model’s own page rather than on its pricing table, Anthropic prints the Claude windows on its models overview, and the Grok figure comes from xAI’s model docs. Where a vendor splits the figure across a model page and a price page, the ledger records which of them each half came from.

Model	Published window	Roughly
GPT-6 Astra (OpenAI)	1,050,000 tokens	several novels
Gemini 3.8 Flash (Google)	1,048,576 tokens	several novels
Claude Sonnet 5 (Anthropic)	1,000,000 tokens	several novels
Grok 4.6 (xAI)	500,000 tokens	a long shelf of documents
Claude Haiku 4.5 (Anthropic)	200,000 tokens	one long novel

The consumer products publish less. On the pages our plans ledger reads, Anthropic states a million token window on Claude’s paid plans, and Google states the same on its AI Pro and Ultra tiers but publishes no window for the cheaper AI Plus, while OpenAI publishes no per plan window for ChatGPT at all. Where a vendor prints nothing, our ledger records nothing rather than an estimate, which is why the same column is populated for one paid tier and empty for the tier below it.

A bigger window is not free

The window is billed by the token, at the input rates explained in our guide to how model pricing works. The headline rate is not the rate that applies at the top of the window, and this is the detail worth learning before you plan a bill around a big number. It sits on OpenAI’s own model page for GPT-6 Astra: a prompt of more than 272K input tokens is billed at twice the input and cache rates, and one and a half times the output, for the full request. The pricing table carries the matching long context columns, where the input line reads $20 per million against the short context $10.

Filling the 1,050,000 token window is by definition a long context request, so one filled window costs about $21 in input alone, before the model writes a word back. Multiplying the advertised window by the advertised price gets you half the bill. The chart above makes the same point from Anthropic’s list: the cheapest Claude is the one with the smallest window. Capacity is the product.

The second cost is subtler. A model can accept a million tokens without attending equally to all of them, and vendors know it: their long context documentation coaches developers on where to place important material. A window size is a limit, not a promise of perfect recall across the whole span. Treat “fits in the window” and “will be used well” as different claims.

What it means for you

For everyday chat, the practical rules are short. Start a fresh conversation for a fresh topic, because an old chat drags its whole history into every request and spends your allowance on it. When a long session starts contradicting itself, the window is the likely reason: ask the model to summarise the conversation, then carry the summary into a new one. And put the material that matters near your question rather than fifty pages up, whatever the advertised size.

When you compare plans or models, read the window the way you now read the price: as a published, dated figure that varies more inside a vendor’s lineup than the marketing admits, and that this site records with its source on the plans and pricing pages. The number changes what a model can do, but only when what you need is genuinely more text at once, and you are willing to pay by the token to use it.