The window is shared between everything you send and the space the answer needs, and the second half is the part budgets forget. A request that adds up to just under the limit does not fit, because the reply has nowhere to go. This lays the whole thing out: the system prompt, the tool definitions, the documents, the conversation so far and this turn, plus a reservation for the response, shown against the model's actual window. It says whether it fits and by how much, and it names the part that is dominating — which is nearly always the history, and nearly always the thing worth fixing first.
Features
Every part of the request sized against the real window
Space reserved for the reply, which is what makes a budget honest
Warns when the reservation exceeds what the model will generate
Warns when history has taken over the window
A stacked bar showing where the window actually goes
Spare capacity, or the overflow, as a number
Model windows carrying the date they were checked
Every field editable, so it fits your request rather than a template
How to use the Context Window Planner
Pick the model you are targeting
Enter the size of each part of the request
Reserve enough tokens for the reply
Read whether it fits, and what is dominating
Example
Input
32,100 sent + 4,000 reserved
Output
Fits, with 963.9K to spare
The reservation is part of what has to fit, not an afterthought.
Common errors & troubleshooting
It says the request does not fit but my numbers are under the limit. — The reply has to fit too. The committed figure is what you send plus what you reserved.
I do not know what my parts cost. — Use the token counter for a real count of a specific piece of text. This tool is for the shape of the budget rather than an exact measurement.
The reservation is flagged as too large. — Every model caps how much it will generate in one response, which is separate from the window. The cap is shown under the field.
History dominates the window. — That is the warning worth acting on. Compaction, context editing or summarising older turns will do more here than any other change.
Frequently asked questions
Why reserve tokens for the reply?
Because the response is generated into the same window. If the input fills it, there is no room to answer — which produces a truncated reply or an error rather than a clear message about the budget.
What is the difference between the window and the output cap?
The window is how much the model can hold at once. The output cap is how much it will generate in a single response, and it is much smaller. Both are checked here, because exceeding either fails.
Are the numbers exact?
The window sizes are as published, with the date they were checked. What you enter for each part is your own estimate, so treat the result as the shape of the budget rather than a guarantee.
What should I do when it does not fit?
Look at which part is largest, which is almost always history or retrieved documents. Retrieving less and summarising earlier turns both help far more than trimming the system prompt.
Does a bigger window solve it?
It buys room, not efficiency. Everything in the window is paid for on every request, so a conversation that fills a large window costs accordingly — which is what the cost tools are for.
Related tools
Prompt Cache Planner — Order the blocks of a request so the cacheable prefix is as long as possible.
Agent Loop Cost Estimator — Cost an agent loop turn by turn, with history re-sent and caching applied.
Text Chunker — Split a document into overlapping chunks that break on real boundaries.
Token Counter — Count tokens for GPT, Claude and Gemini live — exact OpenAI counts plus Claude/Gemini estimates.