How big it is, per November 2026
Every provider publishes its own limits. Take the Claude API as an example: Anthropic’s documentation lists the window per model, and the current generation all has 1 million tokens:
| Model family (Claude) | Context window | Notes |
|---|---|---|
| Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6 | 1M tokens | Default, no beta header, standard pricing |
| Sonnet 5.5, Sonnet 5, Sonnet 4.6 | 1M tokens | Default, standard pricing |
| Haiku 5.5 | 1M tokens | Prompts over 100,000 tokens cost more |
| Fable 5.2, 5.1, 5; Mythos 5.2, 5.1, 5 | 1M tokens | |
| Older models, such as Sonnet 4.5 (deprecated, retiring November 30, 2026) | 200k tokens |
Two related limits from the same page: a single request can produce at most 128k output tokens, and you can send up to 600 images or PDF pages per request (100 on 200k models). Other providers publish their own limits per model; check their model pages, because the numbers change with every release.
What counts toward the window
Everything the model sees in one request, and what it writes back:
- The system prompt
- Every message in the conversation so far, from both sides
- Tool definitions, including those from MCP servers
- Tool results: search output, file contents, API responses
- Documents and images
- The output, including extended thinking
If the input alone is bigger than the window, the request fails with a “prompt is too long” error. Chat interfaces such as claude.ai can manage the window for you, for instance first in, first out, which means the oldest messages quietly fall away.
Some Claude models can see how much of their budget is left (context awareness): Sonnet 5, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5. Newer models such as Opus 4.7 and later, Sonnet 5.5 and Haiku 5.5 instead work with task budgets, in beta.
The common misunderstanding
The window is not memory in the human sense. It is working memory for one answer. Conversation content does not change the model’s weights. An application can store the transcript and load history, instructions or saved memory into later requests, but the window itself keeps nothing between requests.
The second misunderstanding: a bigger window means you can stop choosing. Anthropic’s documentation says the opposite. As token count grows, accuracy and recall degrade, a phenomenon called context rot. A 1M window is room, not a recommendation. The skill is to put the right tokens in it, which is what context engineering is about.
What it means for a team
Because each conversation starts empty, every person on a team rebuilds context from scratch: pasting the same style guide, explaining the same client, describing the same codebase. Claude Code works around this by loading CLAUDE.md files and an auto memory at the start of every session (the first 200 lines or 25KB of the memory). Other AI tools have their own answers, each inside that tool. Personal session history does not transfer automatically; reuse across tools or colleagues needs shared files or a connected knowledge source.
That makes the size of the window a less important question for a team than what goes into it, and where that comes from.
Where Bombos fits
Bombos fills the window from one shared place: it is a plugin that gives every AI tool a team uses, from Claude and ChatGPT to Copilot, Cursor and your own agents, the same memory of team knowledge, learned rules and what is already in mail, documents and repos. Each request gets only what fits the question, so a big window stays a reserve, not a dumping ground. It runs in the cloud or on your own server, and one person can try it first for their own agents.
Related questions
How many words is 1M tokens?
A token is often a piece of a word rather than a whole word. How many words fit depends on the language and the content; code and tables use tokens differently from prose. If it matters, count with your provider's token counter.
Does prompt caching make the context window bigger?
No. Anthropic's documentation is clear: caching changes what you pay, not whether tokens count toward the window.
What happens when I exceed the context window?
If the input alone is larger than the window, the Claude API returns a 400 error, "prompt is too long". Chat apps often handle it for you by summarizing or dropping the oldest messages.
Does the AI remember things between conversations through the context window?
No. Each new conversation starts with an empty window. Anything that carries over, like a memory feature or a CLAUDE.md file, is loaded into the new window again.
