What is Context window?
By Heemang Parmar · Updated July 2026 · Editorial policy
A context window is the maximum amount of text, measured in tokens, that an AI model can process in a single request, covering the system prompt, conversation history, documents, and the model's own response.
Everything counts against the window: the system prompt, the conversation so far, any documents you paste in, and the response the model generates. Once the limit is reached, something has to be dropped or summarized, or the request fails.
For product work, the context window determines how much of your project a model can actually hold in mind. A large window means an agent can read your whole PRD, your existing code, and your design tokens together, and produce output consistent with all of them. A small one forces retrieval tricks and lossy summaries.
A common misconception is that a bigger window solves everything. Models can still weight the middle of a very long context poorly, and stuffing in irrelevant text costs money and can dilute the signal. The craft is giving the model the right context, not the most, which is why RAG exists even for large-window models.
Why does context window matter?
Context windows matter because they set the ceiling on what an AI system can reason about in one pass. Claude's context windows reach 200K tokens, roughly 150,000 words, enough to hold a codebase, a full PRD, and research findings at once; early models with 4K windows could barely hold a long email thread. Window size is a primary factor in which tasks a model can handle at all.
For teams building on AI, the window is also a budget. Every token in it is paid for on every request, and long contexts increase latency, so production systems curate context aggressively: caching stable parts, retrieving only relevant documents, and summarizing history. How well a product manages its context often matters more than which model it calls.
How does context window work?
- 1Everything enters as tokens: The system prompt, conversation history, documents, and tool results are all tokenized and packed together into one input sequence for the model.
- 2The model attends across it: Attention layers relate every token to every other token, which is what makes in-context reasoning possible.
- 3The window fills up: Input plus generated output must fit within the limit, and long agent sessions approach it faster than teams expect.
- 4Overflow gets managed: Applications trim, summarize, or retrieve selectively so the most relevant context survives when space runs out.
Context window vs RAG vs fine-tuning: how do you give a model knowledge?
| Approach | How knowledge gets in | Best for |
|---|---|---|
| Context window | Paste it directly into the request | Project files and documents that fit and change often |
| RAG | Retrieve relevant chunks at query time | Large or growing knowledge bases that cannot fit |
| Fine-tuning | Train behavior into the model's weights | Style and format consistency, not fresh facts |
How is context window used in practice?
One project context, every stage
ProductOS agents share a single project context from idea through deployment, so research, the PRD, and designs stay in view as later agents work. Nothing has to be re-explained between stages.
Your context inside Cursor and Claude
Over MCP, ProductOS makes the PRD, research, and designs available inside Cursor and Claude. Your editor's model gets project context without you pasting documents in by hand.
A builder that reads the whole project
The Fullstack Builder codes inside a live sandbox with the project's shared context available. It can implement features consistent with the PRD and the design decisions made earlier in the pipeline.
Frequently asked questions
What happens when a conversation exceeds the context window?
The application has to make room: it drops the oldest messages, summarizes earlier history, or refuses the request. This is why long chat sessions sometimes 'forget' early details. Agent products manage this automatically with compaction and retrieval, but quality can degrade near the limit, so long-running work benefits from fresh, well-scoped sessions.
Is a bigger context window always better?
No. Larger windows enable tasks that were previously impossible, but models can attend unevenly across very long contexts, and every included token adds cost and latency. Curated, relevant context reliably beats a stuffed window. Bigger windows raise the ceiling; they do not remove the need to choose what goes in.
How big are context windows in 2026?
Frontier models commonly offer 200K tokens, on the order of 150,000 words, and some providers offer windows of a million tokens or more. That is enough for entire codebases or book-length documents in one request. Effective use at those sizes still depends on how well the model attends across long inputs.
Does the model's response count against the context window?
Yes. The window covers the full sequence, so input tokens plus generated output tokens must fit together, and many APIs also enforce a separate maximum output length. A request that arrives near the limit leaves little room for the answer, which is a common cause of truncated responses.
Related terms
- TokenA token is the basic unit of text that AI models read and generate, roughly four characters or three-quarters of an English word; model pricing, context windows, and generation speed are all measured in tokens.
- Large language model (LLM)A large language model (LLM) is an AI model trained on massive text datasets to predict and generate language, powering writing, coding, analysis, and reasoning tools through token-by-token text generation.
- Retrieval-augmented generation (RAG)Retrieval-augmented generation (RAG) is an AI technique that retrieves relevant documents from a knowledge base and inserts them into a language model's context at generation time, so answers are grounded in real, current data instead of training memory.
- PromptA prompt is the instruction given to an AI model that specifies the task, context, constraints, and output format the model should follow, forming the entire interface between your intent and the model's response.