Products, laws and numbers here change fast. Checked October 2026.
In plain words
When we chat with an AI, it does not remember the conversation the way a person does. Every time it replies, it rereads everything in front of it: its instructions, the files we shared, the whole chat so far, and the words it is writing.
The context window is the limit on how much all of that can be. It is measured in tokens, small chunks of text about three quarters of a word each.
When a conversation grows past the limit, the oldest parts drop out of view, or the app quietly squeezes them into a short summary. That is why a very long chat can "forget" what we said at the start.
Windows have grown fast. In 2020, GPT-3 could see about 2,000 tokens, a few pages. As of October 2026, the main models from OpenAI, Anthropic and Google each take about a million tokens, roughly 750,000 words.
Everyday example
A whiteboard in a classroom
A teacher explains a long sum on one whiteboard. As she runs out of space, she wipes the oldest lines on the left to make room for new ones. Students can use only what is still on the board. If step one has been wiped, nobody can check it. An AI's context window is that board: fixed in size, and only what is on it counts.
Where the example breaks: A teacher remembers what she wiped. The model does not: once text leaves the window, it is gone for that reply, unless an app saves notes and puts them back in.
How it actually works
The window is a hard limit set by how the model was built and trained.
Everything shares one budget
System instructions, uploaded documents, earlier turns, results from tools and the reply being written all count against the same number of tokens. A long reply leaves less room for the input.
Why there is a limit at all
In a transformer, every token pays attention to every other token. Double the text and that work roughly quadruples, along with the memory needed. Engineering tricks have pushed limits up, but cost still grows with length.
Big is not the same as sharp
Models do not use every part of a long window equally well. A 2023 Stanford-led study found answers were best when the key fact sat near the start or end, and worse when it was buried in the middle.
- GPT-3, 2020
- 2,048 tokens
- Gemini 1.5, Feb 2024
- 1 million tokens in preview: about an hour of video or 700,000 words, said Google
- Top models, Oct 2026
- About 1 million tokens for the flagship models of OpenAI, Anthropic and Google
- Largest advertised
- Meta's Llama 4 Scout, April 2025: 10 million tokens
The window is short-term attention, not knowledge. What a model learned in training is stored in its parameters, not in the window.
Questions people ask
Why does ChatGPT forget things from earlier in a chat?
Usually because the conversation outgrew the context window, so the earliest messages were dropped or summarised. Starting a fresh chat with a short recap of the key points often works better than continuing a very long one.
What does 128K or 1M context mean?
It is the window size in tokens: 128K is 128,000 tokens and 1M is one million. One English token is about three quarters of a word, so 128K is roughly 96,000 words, about the length of a full novel.
Is a context window the same as memory?
No. The window is what the model can see in one go, like a desk. Memory features are separate notes an app saves between chats and puts back on the desk when needed. Without them, each new chat starts empty.
Is a bigger context window always better?
Not always. Bigger windows let us share whole books or codebases, but long inputs are slower and cost more, and models can miss details buried in the middle. Giving the model only what matters often gives better answers.
Do the AI's own replies count towards the context window?
Yes. The reply is written inside the same window, and so is any hidden thinking a reasoning model does. That is why models also list a separate, smaller limit for how long one reply can be.