What a context window is
A context window is the bounded space available to a model as it processes a request and generates a response. It can contain instructions, your message, earlier conversation turns, attached or retrieved documents, tool results, and generated text.
Providers measure this capacity in tokens. The exact accounting varies. One API may describe a total context limit, while another lists an input limit and a separate maximum output. Either way, the window is about information available now, during inference. It is not a store of facts that the model keeps permanently.
What a knowledge cutoff is
An AI knowledge cutoff is a provider-reported date that summarizes how current the model's trained-in knowledge is. It concerns patterns and facts encoded in the model's weights before deployment.
Treat the date as a guide, not a clean border. Training data mixes sources collected at different times. Coverage also varies by topic. Some providers distinguish the date through which knowledge is considered reliable from a broader training-data cutoff.
The actual distinction
The context window answers: How much information can this request give the model?
The knowledge cutoff answers: How recent is the information the model may know without being given a source?
They are independent properties. A model can have a large context window and an old cutoff. It can also have a smaller window and a newer cutoff. Increasing one does not change the other.
| Dimension | Context window | Knowledge cutoff |
|---|---|---|
| What it limits | Information available in one request | Recency of trained-in knowledge |
| Usually expressed as | A number of tokens | A date |
| Set mainly by | Model and API design | Training data and provider reporting |
| Changes when you paste a document | Available space is consumed | Does not change |
| Can include current information | Yes, if current material is supplied | Not by itself |
| Persists after the request | No; applications must save and resend needed material | Yes, as part of the deployed model's weights |
| Main practical question | Will the needed material fit and remain usable? | Should the model be trusted to know this without a current source? |
A worked example
Suppose a hypothetical model has a 20,000-token context window and a reported knowledge cutoff of December 2025.
You ask about a policy published in March 2026 but provide no source. The request easily fits in the context window. The likely problem is the knowledge cutoff: the model may not have trained-in knowledge of the policy.
You then paste a 2,000-token copy of the policy. The model can now analyze post-cutoff information because the text is inside the current context window. Its knowledge cutoff has not moved. The model has received temporary evidence for this request.
Now suppose the relevant documents total 25,000 tokens. Their publication date is no longer the immediate constraint. The context window is. You must select, split, compress, summarize, or retrieve the most relevant material.
This produces a simple diagnostic:
- The model lacks a recent fact, but you can supply it: use the context window.
- The required material does not fit: manage the context-window budget.
- The model must answer from current facts and a large evidence set: both limits matter.
When each one matters
The context window matters when you work with long documents, extended conversations, large codebases, many tool results, or retrieved evidence. It determines how much can be presented at once and how much room remains for the answer.
The knowledge cutoff matters when a question depends on events, rules, prices, people, research, software, or other facts that change over time. A cutoff should prompt you to seek a current source. It should not be treated as a guarantee that every earlier fact is present or correct.
Both matter in systems that use search or retrieval. Search can fetch current material, but that material still has to enter the context window. The model must then identify and use it correctly. Retrieval works around stale trained-in knowledge; it does not retrain the model or make the window unlimited.
What people confuse
"A bigger window gives the model more knowledge." A bigger window gives you more room to provide information. It does not add facts to the model's weights.
"A new model knows the latest events." Release date and knowledge cutoff are separate. Check the model documentation, then verify time-sensitive claims with current sources.
"Web search removes the cutoff." Search gives the model fresh text to work with. The underlying trained-in knowledge still has a cutoff, and the model can still misread or ignore retrieved evidence.
"If the text fits, the model will use it." Capacity is not the same as attention quality. Long-context research has found that models can miss relevant information, especially when it is buried in the middle of a large input.
"Everything before the cutoff is known." A single date compresses a messy training corpus into one label. Knowledge can be incomplete, uneven, or wrong on either side of that date.
Where to go next
Read What Is a Context Window? for the mechanics of token budgets and long inputs. Read What Is an AI Knowledge Cutoff? for how to interpret cutoff dates and time-sensitive answers.