What Is a Context Window?

A context window is the maximum amount of text — measured in tokens — that a large language model can process in a single request: the prompt, any retrieved documents, the conversation history, and the answer being generated all share it. It is the model's working memory, distinct from the parametric knowledge frozen in its weights.

How big are context windows today?

The range is wide and moving fast. GPT-4o launched with 128,000 tokens; Anthropic's Claude models carry 200,000 as standard; Gemini 1.5 Pro reached 1 million tokens in 2024, and OpenAI's GPT-4.1 (April 2025) matched the million-token class. For scale: a million tokens is roughly 750,000 words — several long novels in one request.

Why doesn't a big window mean your whole site gets read?

Three constraints keep pipelines selective:

• Cost and latency: input tokens are billed and processed linearly, so answer engines serving millions of queries extract main content and cap per-source token budgets rather than ingesting full pages.
• Attention degradation: research on long contexts (the "lost in the middle" finding, Liu et al., 2023) showed models retrieve facts from the start and end of the context far more reliably than the middle. More context can mean less accurate use of any given passage.
• Retrieval design: RAG exists precisely to avoid stuffing windows — select the top passages, not the top sites.

For publishers, the implication is unchanged by window growth: the pipeline keeps fragments, so lead with substance and make every early token count, per the standard GEO guidance.

What changes as windows grow?

Deep Research modes and agentic workflows exploit large windows to hold dozens of full sources simultaneously, enabling multi-document comparison that short windows precluded. That expands opportunity for long-form authoritative content — but position bias persists, so the answer-first discipline holds even when the whole document fits.

Example

Paste a 40-page whitepaper into Gemini and it fits with room to spare; ask a question answered on page 23 and accuracy drops relative to the same fact on page 1 — working memory is not uniform memory. The related mechanics — tokens, attention, extraction — are defined in this glossary.