In this guide · 7 sections
Short answer
The context window is the maximum text, counted in tokens, that an AI model can read and write in one go. Current Claude, GPT and Gemini models list about 1 million tokens, but studies show models are less reliable at finding details in the middle of long inputs, so split long documents.
Tokens: how AI measures text
Language models don't read words; they read tokens, which are chunks of text. A common word may be one token, a long or unusual word several, and punctuation counts too. OpenAI's help article on what tokens are gives rules of thumb for English: one token is about 4 characters or three-quarters of a word, and 100 tokens is about 75 words.
| Text | Rough size in tokens |
|---|---|
| A short email (150 words) | About 200 |
| A 10-page contract (5,000 words) | About 6,700 |
| A 300-page book (90,000 words) | About 120,000 |
| 1 million tokens | Roughly 750,000 English words |
Tokens also matter for cost: API pricing is per million tokens, as our AI API pricing comparison shows.
What the context window is
The context window is the maximum number of tokens a model can work with in one request. It has to hold everything: the system instructions, the whole conversation so far, any files you attached, and the answer the model writes. When a conversation goes past the limit, the app has to drop or summarise older parts, which is when the assistant appears to forget.
Vendor documentation in October 2026 lists large windows across the board:
| Vendor | Models | Context window | Max output |
|---|---|---|---|
| Anthropic | Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 | 1M tokens | 128K tokens |
| OpenAI | GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna | 1.05M tokens | 128K tokens |
| Gemini 4 Argon | 1M tokens | Long outputs (see Google's page) |
Sources: Anthropic's models overview, OpenAI's models page and Google's Gemini models page.
Why AI still misses things in long documents
A bigger window means more text fits, not that every part gets equal attention. The research paper Lost in the Middle (2023) found that performance is often highest when the relevant information is at the beginning or end of the input, and drops significantly when the model must use information in the middle of a long context, even for models built for long contexts.
Later work points the same way. Chroma's July 2025 technical report, Context Rot, tested 18 models and reported that performance varied significantly as input length changed, even on simple tasks. Models have improved since, and vendors publish their own long-context results, but the pattern is the reason to be careful: a long input that fits in the window can still produce an answer that skips the one clause that mattered.
Our guide to AI hallucinations as a business risk explains how to set up checks.
Context vs memory: why it forgets between chats
The context window covers one conversation or request. When you start a new chat, the model starts with an empty context apart from its instructions. Some chat apps add a separate memory feature that saves selected facts or preferences and inserts them into future chats. That is useful, but it is a short list of saved notes, not a record of everything you discussed.
Context window
- Everything in the current conversation and attached files
- Fixed maximum size in tokens
- Cleared when you start a new chat
Best for: Working through one document or task.
Memory or projects
- Selected facts, preferences or files saved by the app
- Added to new chats automatically
- Limited and sometimes wrong; review and edit it
Best for: Recurring preferences and reference files.
Practical tips for working with long documents
These habits make answers on long material more reliable, whichever assistant you use:
Ask focused questions
"List every termination clause with its clause number" works better than "Summarise this contract".
Split long documents by section
Work through a 200-page tender a few sections at a time, then combine the answers yourself.
Put the question after the document
Paste the material first and your question at the end, so the instruction sits where models attend well.
Ask for references
Ask the model to quote or cite the page, clause or heading behind each point, then spot-check a few.
Start a fresh chat for a new task
Long, wandering conversations fill the window with old material. A clean chat with a short summary of what matters works better.
Use retrieval for large libraries
For hundreds of documents, a set-up that finds the relevant passages first and gives only those to the model beats pasting everything in.
The last approach is called retrieval-augmented generation; our guide to RAG and chatting with your documents explains it. Published long-context benchmarks are covered in our guide to LLM benchmarks.
Do it yourself or get help?
For occasional long documents, the tips above are enough with any mainstream assistant. You can run a quick test yourself: take a document you know well, ask five specific questions about details from the beginning, middle and end, and see which ones the assistant gets right.
If your business needs an assistant that answers from a large set of your own documents, such as policies, product manuals or past proposals, our AI Website Assistant (S$790) sets one up with retrieval over your content, source references in each answer, and handover to a person when it isn't sure.
Frequently asked questions
It is the maximum amount of text, measured in tokens, that the model can consider in one request, including instructions, the conversation so far, attached files and its own answer. As of October 2026, the API documentation for current Claude, GPT-6 and Gemini 4 models lists windows of about one million tokens.
Using OpenAI's rule of thumb for English that 100 tokens is about 75 words, one million tokens is roughly 750,000 words. Other languages, code, tables and unusual formatting use more tokens per word, so the real figure for your documents may be lower.
Fitting in the window doesn't mean equal attention to every part. Research such as the 2023 Lost in the Middle paper found models perform best when relevant information is near the beginning or end of the input and worse when it sits in the middle. Splitting documents and asking focused questions helps.
Each new chat starts with a fresh context window. Unless the app's memory or project feature saved a fact and adds it to the new chat, the model has no access to your earlier conversation. Memory features store selected notes, not full histories.
Not always. A larger window lets you include more material, but long inputs cost more tokens, take longer and can reduce accuracy on details. For large document collections, retrieving only the relevant passages and giving those to the model usually gives better, cheaper answers.