The context window is the model's working memory: everything Claude can see at once in a single request — your instructions, the documents you attach, the conversation so far, and the response it is writing. As of July 2026, Claude Opus 4.8 and Claude Sonnet 5 offer a 1M-token context window, while Claude Haiku 4.5 offers 200K tokens. Those are big numbers, and the point of this article is to make them concrete: what fits, what it costs, and when a huge window is actually the right tool.
Translating tokens into documents
A token is the unit models read and bill in — roughly a short word or a piece of a longer one. As a working approximation, a page of ordinary business text is on the order of several hundred tokens, so a million tokens comfortably holds material on the scale of a long due-diligence binder, a substantial codebase slice, or months of correspondence on a matter — in one request, with no retrieval machinery in between. Haiku 4.5's 200K window is smaller but still holds what most people would call a very large document set. (For the fuller primer on the unit itself, see Tokens 101.)
The qualitative shift matters more than the arithmetic: below a certain size, you must build a retrieval system that selects relevant excerpts to show the model; above it, for many workloads, you can simply include everything relevant and let the model find what matters. That removes a whole category of engineering — and a whole category of retrieval-missed-the-key-paragraph failures.
What a full window costs
Input tokens are billed per million, so filling a 1M window has a straightforward price: at list rates, one maximally full request to Opus 4.8 is about $5.00 of input, and to Sonnet 5 about $3.00 ($2.00 at the introductory rate through August 31, 2026) — before output tokens, which are billed separately and at higher rates. One such request is cheap compared to an hour of an analyst's time. The same request repeated thousands of times a day is a real line item.
That is why prompt caching is the essential companion feature. If many requests share the same large prefix — the same policy manual, the same contract set, the same codebase — caching lets the platform reuse that processed prefix across requests instead of re-billing it at full rate every time. Prompt caching is available on all four platforms, and for large-context workloads it is routinely the single biggest cost lever you have. The details are in the prompt caching article.
Big context vs. RAG: not either/or
Large windows have not made retrieval-augmented generation (RAG) obsolete; they have changed where the boundary sits. If your knowledge source is bounded and stable — a policy manual, a contract portfolio for one deal, a product's documentation — loading it directly into context is simpler and often more accurate than building retrieval. If your source is effectively unbounded or fast-changing — the whole corporate wiki, years of tickets, a document store growing daily — you still need retrieval to select what goes into the window, because no window holds everything and you would not want to pay to send it anyway. Many production systems combine both: retrieval narrows millions of documents to the relevant few dozen, and the large window means "few dozen" can be sent whole rather than as fragments. See RAG basics for that side of the design.
Practical habits for large-context work
A few habits keep big-window applications healthy. Structure your input: label documents clearly and put instructions where they are easy to find, because a well-organized prompt is easier for the model to use faithfully than an undifferentiated dump. Put stable content first and variable content last, which is also the layout that makes prompt caching effective. Track token usage per request in your logs from day one, so cost conversations are grounded in data. And remember the model-tier difference when routing: a workload built around 1M-token requests belongs on Opus 4.8 or Sonnet 5, while Haiku 4.5's 200K window pairs with the high-volume, smaller-context tasks it is priced for.
Where to go next
Cost and context are inseparable at this scale — continue with prompt caching: the single biggest cost lever, or RAG basics for when retrieval still earns its keep. The model lineup summarizes context sizes and pricing side by side.