What Is a Context Window?

The context window is all the text an AI model can see at once — your prompt, its response, and everything in between.

Every time you send a message, the model doesn't re-read the internet. It reads a fixed-size block of text. Think of it like a whiteboard: there's a maximum amount that fits, and once it's full, older content must make room.

Key distinction:

Training data — everything the model learned during development. Permanent.
Context window — the text the model actively references right now. Resets every new conversation.

Current sizes:

1M tokens — Claude Opus 4.7, Opus 4.6, Sonnet 4.6
200K tokens — Claude Haiku, most GPT models
128K tokens — GPT-4o

That sounds like a lot, but tokens add up fast. A long coding session can burn through hundreds of thousands of tokens in minutes.

How Context Builds Up

Each turn in a conversation adds tokens — previous turns are preserved completely.

Context grows progressively:

Turn 1: [system + your msg + AI reply]
Turn 2: [system + your msg + AI reply + your msg + AI reply]
Turn 3: [system + ... + your msg + AI reply]

What counts toward the limit:

System instructions and tool definitions
Your previous messages and the model's previous responses
Files, images, or code you've attached
The response currently being generated

Why More Context Isn't Always Better

Context rot — as the window fills up, accuracy and recall can degrade.

More context ≠ better answers. Models can lose track of details buried deep in a long conversation. This isn't a bug — it's a known limitation of how attention works in transformers.

Signs your context is getting crowded:

The model forgets instructions from 20 turns ago
It contradicts something it said earlier
Answers become generic instead of specific

The fix isn't always a bigger window. Often it's a cleaner one.

How Models Manage Context

Modern models use several techniques to keep long conversations going without hitting the wall.

Compaction — When a conversation grows too long, the server automatically summarizes older turns. Some detail from early turns gets compressed away.

Extended thinking — Thinking blocks are generated during output but automatically stripped from subsequent turns. Your next message sees the final answer, not the internal reasoning.

Context awareness — Claude 4.5+ tracks how much context remains. As the window fills, the model adjusts — shorter responses, more concise reasoning.

Prompt caching — Anthropic caches your prompt prefix for up to 5 minutes. Reuse within that window costs 90% less. One-hour caching available for frequent prompts.

Tips for Managing Your Context

Being intentional about what's in your context window beats stuffing more in.

Start fresh when the topic shifts. A new conversation gives you a clean window with no leftover instructions.

Be selective about attachments. Pasting an entire 500-line file when you need one function wastes tokens. Paste the relevant section.

Put important instructions at the edges. Models pay more attention to the beginning and end of the context. Don't bury critical rules in the middle.

Use system prompts for persistent rules. Instead of repeating "respond in Python" every turn, put it in the system prompt.

Try it

Context Windows — Anthropic Docs

Prompt Caching — reduce cost and latency for repeated context