On this page
Definition
The context window is the token budget an LLM can read and write within a single request, covering the prompt, retrieved data, previous turns and the response itself.
Simple explanation
Think of the model's context window as its short-term memory. Bigger windows mean the model can 'see' more at once — a longer document, more chat history or more retrieved data.
When a conversation gets long, older messages may be pushed out. When you paste a huge document, only what fits will be processed.
Why it matters
The context window determines what an AI system can actually do in one shot. Long-context models power document analysis, code understanding and long conversations without losing thread.
It also determines cost — longer inputs cost more and take longer.
How it works
- 1TokeniseYour input is split into tokens (roughly 3-4 characters each on average).
- 2BudgetAll tokens — prompt, history, retrieved docs, reply — must fit within the model's limit.
- 3Truncate or summariseSystems drop or compress older content to stay within budget.
Real examples
Products named for illustration only. Inclusion is not an endorsement.
- Gemini 2.5Up to 2M-token windows suited to very long documents.
- Claude Sonnet 4.5Long-context reasoning across code repositories.
- GPT-5Large windows tuned for reasoning tasks.
Advantages
- Bigger windows unlock document-scale tasks.
- Long conversations feel coherent.
- More retrieved context can be shown at once.
Limitations
- Cost and latency rise with input length.
- Models can still miss detail in the middle of very long inputs.
- Bigger is not always better — precision often beats volume.
Common misunderstandings
- ClaimA larger context window means the AI remembers between chats.RealityIt does not — memory across sessions is a separate feature.
- ClaimMore context always means better answers.RealityBeyond a point, irrelevant context reduces answer quality.
Frequently asked questions
How many words is one million tokens?
Roughly 700,000 English words.
Does context reset when I start a new chat?
Yes, unless the product has an explicit memory feature.
Is a bigger context always better value?
No — for narrow tasks a smaller, cheaper model often wins.
The Tool Money Lab perspective
Long-context is one of the most under-marketed capability jumps of the last two years. Analyst-grade document work that took hours can now be first-drafted in minutes.
Conclusion
Context windows define what a model can see. Bigger windows enable bigger tasks — but only if the input is signal, not noise.
Design prompts and RAG pipelines to feed the model what it needs, not everything you have.