Nothing beyond the window is read. When input alone exceeds the budget, the tool must drop, summarize, or refuse part of it.
The exact behavior varies by model and tool: some drop the oldest content, some compress it, some report an error. Measure it on the tool in use.
Overflow policy | What the model sees | Risk |
|---|---|---|
Drop oldest | Recent content only | Early instructions and constraints vanish silently. |
Summarize oldest | A compressed gist | Details needed later are gone. |
Report an error | Nothing until the input shrinks | The run stops instead of degrading. |
Generation also stops when the budget runs out. A reply cut off mid-answer usually means the output reserve was too small, not that the model failed.
Long histories shrink every future answer. Each turn banks more input tokens that later turns cannot reuse for output.
The calc block reuses the EXAMPLE plan of 2150 tokens against an EXAMPLE budget of 2000 tokens. Cost uses an EXAMPLE price of 3 units per 1 million input tokens.
EXAMPLE budget: 2000 tokens
planned use: 2150 tokens
over by 2150 - 2000 = 150 tokens
EXAMPLE input price: 3 units per 1 million input tokens
actual planned input = 100 + 750 + 800 = 1650 tokens
input cost if sent and billed = 1650 x 3 / 1000000 = 0.00495 units
the 500-token answer reserve is capacity, not billed input
actual generated output is priced separatelyThe earliest content: system rules or early constraints. Dropping oldest removes the head of the conversation first.
No. The planned input is 1650 tokens, costing 0.00495 example units if sent and billed. Reserved capacity is not token usage; generated output is accounted for separately.