Every conversation you have with an LLM exists within invisible boundaries. The model can only “see” a limited amount of text at once, and everything beyond that boundary simply does not exist to it. This boundary is called the context window, and understanding it changes how you work with any large language model.
The context window determines what information the model can access when generating a response. It includes your current prompt, any previous messages in the conversation, and any documents or examples you provide. Once your conversation exceeds this limit, older content gets pushed out and becomes invisible to the model.
For anyone using LLMs regularly, context window limits affect everything from document analysis to long conversations. Knowing how context windows work helps you structure requests more effectively. It also helps you avoid the frustrating moment when a model “forgets” something you told it earlier.
Key Takeaways
What the Context Window Actually Is
Context window: The maximum amount of text an LLM can process in a single conversation turn, measured in tokens. It includes everything the model reads (your prompt, previous messages, system instructions) and everything it writes back.
Think of the context window as a fixed-size workspace. Everything the model needs to read and everything it produces must fit inside that workspace. Once the space fills up, the model cannot take in more information without losing something from the beginning.
This workspace is measured in tokens, not words. A token is the smallest unit of text an LLM processes. In English, one token roughly equals three-quarters of a word.
A 200,000-token context window can hold approximately 150,000 words, or about 500 pages of text.
The context window is shared between input and output. If a model has a 200,000-token context window and you send 190,000 tokens of input, the model only has 10,000 tokens left for its response.
This is why some models specify a separate maximum output length. Claude Opus 4.6, for example, has a 1-million-token context window but limits individual outputs to 128,000 tokens.
Why Context Windows Exist
Context windows exist because of how large language models work internally. LLMs use a mechanism called self-attention that compares every token against every other token in the input. The computational cost of this process grows rapidly as the token count increases.
Doubling the context window roughly quadruples the computation required. Context windows stayed relatively small, around 4,000 tokens, for years because of this cost. Specialized engineering techniques like sparse attention and sliding window mechanisms eventually made longer contexts practical.
Today’s frontier models have pushed context windows dramatically larger. But the fundamental trade-off remains: bigger windows allow more information at the cost of more processing power, more time, and more money.
How Context Windows Have Grown
The growth in context window sizes over the past few years has been dramatic. GPT-3 launched with a 4,096-token window in 2020. By late 2023, models like GPT-4 Turbo reached 128,000 tokens.
In 2025 and 2026, several models crossed the 1-million-token barrier. Advances in efficient attention algorithms, better hardware, and techniques like ring attention made this possible. These methods let models process longer sequences without the cost scaling as sharply as it once did.
This growth matters because it expanded what LLMs can do in a single interaction. A 4,000-token window could barely hold a few pages of text. A 1-million-token window can hold an entire book, a full codebase, or months of conversation history.
How It Manifests in Practice
The context window affects your experience with an LLM in three main ways: conversation memory, document processing, and output quality.
Conversation Memory
Every message you send and every response the model generates accumulates inside the context window. Early in a conversation, the model remembers everything. As the conversation grows, it approaches the context limit.
When the limit is reached, most chat interfaces handle this silently. They remove the oldest messages from what the model can see. This means the model gradually forgets what you discussed at the start.
You might notice the model repeating something you already told it, contradicting an earlier answer, or forgetting a name you mentioned ten messages ago.
Document Processing
If you paste a document into an LLM, the entire document must fit within the context window. Your instructions and the model’s response also need room inside that same limit.
A 10,000-word report uses roughly 13,000 tokens. A 100-page contract might use 80,000 tokens or more.
Models with smaller context windows simply cannot process long documents in one pass. You would need to split the document into chunks and process each chunk separately. This loses the ability to reason across the whole document at once.
For example, say you ask a model with a 128,000-token window to review a 200,000-token codebase. It can only see part of the code at a time. A bug that spans two files might be invisible if those files land in different chunks.
Output Quality Over Distance
Research shows that models can struggle with information placement. Even when text fits within the context window, the model may underuse information from the middle of very long inputs.
This is sometimes called the “lost in the middle” effect. Information at the beginning and end of the input tends to be used more reliably than information buried in the center.
This means filling a context window to capacity is not always the best strategy. Placing your most important instructions at the start or end of the prompt can produce better results than scattering them throughout a long document.
Context Window Sizes by Model (February 2026)
The following table shows current context window sizes for major LLM models, based on official specifications as of February 2026.
| Model | Provider | Context Window | Max Output | API Input Cost (per 1M tokens) |
|---|---|---|---|---|
| GPT-5.2 | OpenAI | 400,000 | 128,000 | $1.75 |
| GPT-5 | OpenAI | 400,000 | – | $1.25 |
| GPT-5 nano | OpenAI | 400,000 | 128,000 | $0.05 |
| Claude Opus 4.6 | Anthropic | 1,000,000 | 128,000 | $5.00 |
| Claude Sonnet 4.5 | Anthropic | 200,000 | – | $3.00 |
| Claude Haiku 4.5 | Anthropic | 200,000 | – | $1.00 |
| Gemini 3.1 Pro | 1,000,000 | 64,000 | $2.00 | |
| Gemini 2.5 Pro | 1,000,000 | – | $1.25 | |
| Gemini 2.5 Flash | 1,000,000 | – | $0.15 |
Sources: OpenAI pricing, Claude pricing, Google AI pricing. Prices reflect standard API rates. Some providers offer discounts for cached or batched input.
Two patterns stand out here. First, Google’s Gemini models lead in raw context size, offering 1 million tokens across their entire lineup. Second, larger context windows do not always mean higher prices.
Gemini 2.5 Flash provides 1 million tokens of context at $0.15 per million input tokens, while GPT-5.2 offers 400,000 tokens at $1.75.
The max output column matters too. Even with a massive context window, the model can only generate a limited response. If you need the model to rewrite an entire book chapter, verify that its max output length supports that.
Strengths and Limitations
When Large Context Windows Help
Large context windows are genuinely useful for specific tasks. Analyzing long legal contracts, reviewing entire codebases, and comparing multiple documents side by side all benefit from a large context. Extended conversations that span dozens of messages also depend on it.
For research tasks, a large context window means you can paste several papers and ask the model to find contradictions or synthesize findings. This is something that AI-assisted research depends on.
A practical example: comparing three vendor proposals that total 40,000 words. With a large enough context window, you can load all three at once. Then ask the model to identify pricing differences, missing deliverables, or conflicting terms. Splitting them across separate prompts would make cross-document comparison impossible.
When They do Not Help
A larger context window does not make the model smarter. It simply lets the model read more text at once. A poorly written prompt performs just as poorly in a 1-million-token window as it does in a 200,000-token window.
There are also practical limits. Processing more tokens takes longer. A query using 500,000 tokens will be noticeably slower than one using 5,000 tokens.
Costs scale linearly with token count. A long-context query follows the same token-based pricing structure as any other API call, just with a larger bill.
Filling a context window to capacity often degrades output quality. The “lost in the middle” effect means the model may ignore information buried in the center of very long inputs. Use the minimum context needed for your task rather than loading everything at once.
The “lost in the middle” problem also applies. Stuffing a context window full of loosely relevant information can actually degrade performance compared to sending a focused, shorter prompt.
Strategies for Working Within Limits
When your content exceeds the context window, or when you want to use it more efficiently, consider these approaches:
- Summarize before sending. Condense long documents into key points first, then ask the model to work with the summary.
- Chunk strategically. Split documents at natural boundaries (chapters, sections) rather than arbitrary cut points.
- Put key instructions first and last. Take advantage of the model’s stronger attention to the beginning and end of the input.
- Use retrieval-augmented generation (RAG). Instead of pasting entire documents, use a system that retrieves only the most relevant passages for each query.
Common Misunderstandings
“A Bigger Context Window Means a Better Model”
Context window size measures capacity, not intelligence. A model with a 200,000-token window might produce higher-quality analysis than a model with 1 million tokens if its underlying architecture is more capable. When choosing an LLM, treat context window size as one factor among many, not the deciding one.
“The Model Remembers Everything in the Window Perfectly”
Models process the entire context window, but they do not give equal weight to every part. As noted earlier, information in the middle of long inputs can receive less attention than information at the boundaries.
This is not the same as hallucination, where the model invents information. It is a limitation of how attention mechanisms distribute focus across long sequences.
“Context Window and Memory are the Same Thing”
LLMs have no persistent memory between separate conversations. The context window is temporary. Once a conversation ends, everything in it disappears.
Some platforms add memory features on top, but the underlying model starts fresh each time. The context window is better understood as a temporary workspace rather than long-term storage.
“Tokens and Words are the Same Thing”
This is a common source of confusion. A model with a 200,000-token context window does not hold 200,000 words. In English, the ratio is roughly 1 token per 0.75 words, so 200,000 tokens is closer to 150,000 words.
In other languages, especially those with complex scripts, one word might require several tokens. This reduces the effective capacity even further. Understanding how tokenization works helps avoid surprises.
Conclusion
The context window defines the boundaries of what an LLM can consider in a single interaction. It is measured in tokens, shared between input and output, and varies significantly across models.
Understanding this constraint helps you write better prompts and choose the right model. It also saves you from the frustrating experience of a model forgetting what you told it five messages ago.
If you are just getting started, knowing the context window sits alongside understanding how LLMs process language and recognizing their limitations as a foundational skill. For practical guidance on picking a model that fits your needs, including context window size, the right LLM comparison can help you weigh the trade-offs.
Frequently Asked Questions
