What Is a Context Window? (2026)

Windows have grown enormously. As of June 2026, the top-tier models offer 1M-token context: Anthropic includes a 1M-token window at standard pricing on Claude Opus 4.6 and newer, Sonnet 4.6, and Fable 5 (see Anthropic pricing and the Claude API pricing detail). Google's Gemini line is also known for very large context, documented on the Gemini pricing page, where the Gemini 3.1 Pro preview tier is priced for context up to 200k tokens.

A 1M-token window is roughly 750,000 words — on the order of several long books at once. That is enough to drop entire codebases, long contracts, or months of transcripts into a single prompt without retrieval. Exact limits vary by model and change often, so check the provider's live docs before you architect around a specific number.

A large window is a budget, not a free resource — you pay per input token for everything you put in it. Filling a 1M-token window on a model priced around $3–$5 per million input tokens costs several dollars per call before the model writes a single word of output. Do that on every request and costs add up fast.

Some providers also adjust pricing at very large context. Note that on Gemini, the published preview rate for Gemini 3.1 Pro applies to context up to 200k tokens (see Gemini pricing); always confirm the long-context tier before relying on it. The practical takeaway: a bigger window lets you include more, it does not make including more cheaper.

Prompt caching is the main lever here. If a big chunk of your context is stable across calls — a long system prompt, a reference document, a codebase — cache it. On Anthropic, cache reads are about 10% of the base input price, so a reused 1M-token prefix becomes far cheaper after the first call (details on the Claude API pricing page). To model the dollars for your own setup, use the AI prompt cost calculator.

Not automatically. Having room to include more is not the same as the model using all of it well. Models can still lose track of details buried in the middle of a very long context, and irrelevant filler can dilute attention and add noise. More context can mean a worse answer if most of it is off-topic.

The skill is curation, not maximization. Put the most relevant, highest-signal material in the window and leave the rest out. This is precisely the problem retrieval-augmented generation solves — fetching only the passages that matter instead of stuffing everything in. See what is RAG.

Budget it deliberately. Estimate the tokens for each part of your prompt and leave clear headroom for the output — remember the model's answer competes for the same budget.

Lead with the instructions and the most important material, and place key constraints where they are least likely to be lost. For long-running chats, summarize earlier turns instead of resending raw history. When your knowledge base dwarfs the window, retrieve the relevant chunks rather than including everything. And cache anything stable to keep repeated large prompts affordable.