Tools
Context budget
A context window is shared between everything you send and everything the model can reply with. Fill in each part, as a token count or by pasting the text, and see the split as a stacked bar, what is left for the answer, and what overflows. It is usually the history and the tool definitions that take the room, not the user's question.
Windows as published on 2026-09-10. Providers that do not publish one next to their prices are left out; use "Custom window" for those.
Room left for the answer. The window is shared: every input token you add is one the model cannot use to reply.
Input
20,820
tokens
Window used
12.4 %
24,820 / 200,000
Free
175,180
tokens left
Largest part
Conversation history
58 % of input
Input composition
175,180 tokens spare. Bear in mind that filling a window is not free: long contexts slow the first token down and cost the same per token wherever they sit.
Stable across calls, so the first thing worth putting behind a cache breakpoint.
Sent on every call whether the model uses them or not. Twenty verbose tools runs into thousands of tokens.
Resent in full each turn because the API is stateless. The term that grows without anyone noticing.
Chunk count times chunk size, which is why the chunker settings show up on your bill.
The actual question. Almost always the smallest part, which is the point of this chart.