Context window

The maximum token budget a model request can use for input and, depending on the API, generated output or reserved reasoning tokens.

Instructions, messages, retrieved documents, tool definitions, and tool results consume context tokens. Providers specify how the output budget interacts with that limit, and products may compact, retrieve, cache, summarize, or truncate material before the model call. Tokens do not map to a fixed fraction of a word across languages, code, images, and tokenizers. The advertised limit states capacity, not equal attention or reliable use at every position.

Builder example

A workflow can exceed a provider limit, lose older material through product compaction, or remain under the limit while irrelevant context reduces task performance. Track the assembled request, source selection, and output reserve, then test the longest representative cases instead of relying on the maximum number alone.

You paste an entire project folder into a chat and ask a specific question. The model misses the answer because it was on page 47 of 100.

Give the model the specific file or section it needs, not the entire folder. Less clutter, better answers.

Common confusion: A larger window allows more tokens but does not create cross-session memory and does not guarantee that adding material helps. The product may load persistent state separately into each request.