What is a context window? How much an AI can read

What a context window means for AI models

A context window is all the text a model can refer to while it writes a reply, including the reply as it is being written. People also call it the context length. The "model" here is the engine inside ChatGPT, Claude or Gemini that actually produces the answer.

Context window at a glance

Also called	Context length
What it is	The text a model can see while writing one reply
Unit	Tokens (small chunks of text)
Closest analogy	Working memory or short-term memory
Not the same as	Training data, which never sits in the window

Working memory, not training data

The context window is separate from everything the model learned in training. A model is built from a huge pile of text, but that pile is not loaded into each conversation.

Anthropic describes the window as the model's "working memory," and Google compares it to short-term memory. You can only hold so much in your head at once, and a model can only look at so much text for any one reply.

Measured in tokens, not words

Window sizes are given in tokens, the small chunks a model splits text into before reading it. A short common word is often one token, while a long or unusual word can be split into several.

How much text fits in a token depends on the model. Google's rule of thumb for Gemini is about 4 characters per token, so 100 tokens comes to roughly 60 to 80 English words.

Everything in the chat counts toward the limit

What you type is only a small part of what sits in the context window. The app's own instructions, every earlier message, attached files and the reply the model is writing all draw on the same budget.

What takes up space in the context window

What	Example
System instructions	The setup text an app sends before your first message
Chat history	Every earlier question and answer
Attachments	Images, PDFs, documents
Tool results	Search results, code output
Your new message	The question you just typed
The reply	The answer being written, plus any visible reasoning

Instructions, files and the reply all use space

Anthropic lists four things that count toward the window. Some of them never appear on your screen.

• The system prompt
• Every message, including tool results, images and documents
• The definitions of any tools the model can call
• The model's own output, including its extended thinking

Claude Code, Anthropic's coding tool, sends a large block of setup text along with your very first message. We break down that overhead in "Claude Code Token Usage: 33K vs OpenCode's 7K."

Each turn adds to the history

Every time you send a message, the model reads the whole conversation again, so the history keeps growing. Your second question arrives with the first exchange attached, and your third arrives with the first two.

Earlier turns are kept in full, not trimmed. Even short messages add up, and a long enough chat will eventually reach the limit.

What happens when the window fills up

What happens at the limit depends on whether you are using a chat app or calling the model through an API. Chat apps tend to quietly let old material go, while APIs refuse or stop.

What happens at the limit

Where	What happens
Chat apps such as claude.ai	The oldest messages may drop out
API, input already too long	The request fails with "prompt is too long"
API, reply hits the limit	The reply stops where it is
Long conversations	Compaction summarizes older turns

Chat apps drop the oldest messages

Chat apps such as claude.ai can handle a full window by dropping the oldest messages first. From then on, each new exchange pushes the earliest one out.

Once a message has dropped out, the model can no longer see it. That is why a chatbot can seem to forget instructions you gave at the start of a long conversation. If something matters, repeat it later in the chat.

APIs return an error instead of dropping text

An API is the interface programs use to call a model, and it does not drop anything for you. On Anthropic's API, a request whose input alone is over the limit fails on every model with a "prompt is too long" error.

If the input fits but the reply runs into the limit, Claude 4.5 and newer models stop at that point. Developers count tokens before sending to avoid both cases.

Compaction summarizes older turns

For conversations that run up against the limit, compaction replaces earlier turns with a summary so the chat can keep going. Anthropic runs this on its servers for long-running conversations and agents.

A summary keeps the gist, not the exact wording. If you need numbers or conditions kept word for word, restate them in a recent message.

Context windows grew from 8,000 to 1 million tokens

Context windows have grown more than a hundredfold in a few years. Google says early generative models could take only 8,000 tokens at a time, and the newest models from Google and Anthropic now take 1 million.

How context windows grew (per Google)

Tokens a model could take at once (thousands)

From early generative models to Gemini. Source: Google, Long context.

8Early

32Larger

128Larger still

1,000Gemini

02505007501,000

Gemini was the first model to take 1 million tokens

Google says Gemini was the first model able to accept 1 million tokens. Before that, windows had grown from 8,000 tokens to 32,000 and then 128,000.

With smaller windows, Google notes, developers often have to drop old messages, summarize them or set up retrieval, which fetches only the relevant passages. A million tokens is enough to hand over a whole set of documents up front.

The newest Claude models have 1 million tokens

As of October 2026, Anthropic's newest models, including Claude Opus 5.5 and Claude Sonnet 5.5, have a 1-million-token context window. Older models such as Claude Sonnet 4.5 have 200,000.

Limits also vary by product, not just by model. OpenAI's Codex, for example, uses a smaller window than its models support, as we explain in "Codex context window cut to 272K: why that number." Each vendor's model list is the place to check current figures.

1 million tokens is about 50,000 lines of code

By Google's count, 1 million tokens is about 50,000 lines of code at 80 characters a line, or eight average-length English novels. Those figures assume English text, and how many tokens other languages take depends on the model's tokenizer.

A bigger context window isn't always better

A larger window lets you include more, but filling it does not make answers better. Accuracy drops as the window fills, and information in the middle of a long input is the most likely to be missed.

Working with long inputs

Problem	What to do
Accuracy drops as more text goes in	Include only what is relevant
The middle of long inputs gets missed	Put key facts at the start or end
The question gets buried	Ask your question after the material

Context rot lowers accuracy as the window fills

Anthropic says that as token count grows, a model's accuracy and recall degrade, which it calls context rot. More room in the window does not mean the model uses all of it equally well.

That makes choosing what to include as important as how much fits. Pasting in loosely related material can crowd out the part you actually need the model to read.

Models miss information in the middle

Models tend to use information at the start and end of a long input well and the middle much less well. Liu and colleagues reported this in their 2023 paper "Lost in the Middle," which tested multi-document question answering and key-value lookups.

The drop showed up even in models built for long inputs. Put the facts and conditions that matter most near the top or the bottom, not buried halfway through.

Put only what matters in, and ask last

With long inputs, include only the relevant material and put your question after it. Google says that when the context is long, placing the query at the end of the prompt usually gives better results.

To summarize a 50-page report, for example, paste the chapters that matter and then write "List the three main points of this report" underneath. Deciding what goes into the window, and in what order, is known as context engineering; we cover Anthropic's guidance in "Context Engineering: Claude Code's Prompt Cut 80%+."

Bigger windows help only if the model reads them well

A context window is the text a model can see while writing one reply, counted in tokens. Your prompt, the app's instructions, the chat history, attachments and the reply all share it, and when it fills, chat apps drop old messages while APIs return an error.

Windows now reach 1 million tokens on the newest Claude models, but accuracy still falls as they fill and the middle of long inputs gets missed. Include only what matters and ask your question last.