Context Window

A context window is the maximum amount of text, measured in tokens, that a language model can take into account in a single request. It covers everything the model sees: system instructions, the conversation so far, retrieved documents, tool results and, depending on the provider, the response it is generating. Anything outside the window does not exist for the model during that request.

Window sizes have grown by more than two orders of magnitude. ChatGPT launched in November 2022 on a GPT-3.5 model with about 4,000 tokens of context. As of October 2026, the current models from the three largest labs all sit around 1 million tokens:

• Anthropic's Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 have a 1 million token window and up to 128,000 output tokens. Anthropic notes that 1 million tokens is roughly 555,000 English words on its current tokenizer.
• OpenAI's model list gives GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna a 1.05 million token window and 128,000 output tokens.
• Google describes Gemini 4 Argon as having a unified 1 million token window with long outputs of up to 1 million tokens.

A large window does not mean the model uses all of it equally well. Research such as 'Lost in the Middle' (2023) and Chroma's 'Context Rot' study (2025) found that accuracy drops as inputs get longer and that information buried in the middle of long inputs is recalled less reliably. Long inputs also cost more and take longer to process. That is why context engineering, choosing the smallest set of relevant material for each step, matters even with million-token windows.

When a conversation or agent run approaches the limit, applications handle it in a few ways: they drop the oldest messages, summarize earlier turns into a shorter version (often called compaction), or store notes outside the window and reload them when needed. Chat products usually do this automatically, which is why a long conversation can 'forget' details from the start.

For AI search and GEO, the context window explains a step that is easy to overlook. An answer engine does not read the whole web for each question. It runs searches, fetches a set of pages, and places selected passages from them into the model's context, then generates the answer from that material (retrieval-augmented generation). Some systems filter aggressively before anything reaches the model: Anthropic's newer web search tool versions have Claude write code that filters search results first, so only relevant content enters the window. Pages compete for that limited space with every other retrieved source.

Practical implications for content: put the direct answer to a question at the start of the relevant section, keep each section understandable on its own so a single extracted passage still makes sense (content chunking), state key facts such as prices, dates and specifications in plain text, and keep pages focused on one topic so the relevant passage is easy to find. These habits help whether a model has a 128,000 or a 1 million token window, because the limiting factor is usually the engine's selection, not raw capacity.