Every turn re-sends the whole context at full price
Input tokens per turn, drawn cumulatively across one minute · Tier 1 allows 30k input tokens/min
30k input tokens / min
Turn 1~16k input tokens
16k
Turn 2~20k input tokens
36k> 30k/min
429
re-sent, byte-identical
Turn 3~23k input tokens
59k
re-sent, byte-identical
0 10k 20k 30k 40k 50k 60k cumulative input tokens within one minute →
system prompt (87 lines) 13 tool schemas workspace file listing message history new message re-sent, byte-identical ITPM limit
Only the last slice of each request is new. The rest is replay billed at full rate — at ~15–25k tokens a turn against a 30k/min budget, turns 1 and 2 together cross the line and turn 2 draws the 429.
View the same data as a table
SegmentTurn 1Turn 2Turn 3
system prompt (87 lines)~3k~3k †~3k †
13 tool schemas~6k~6k †~6k †
workspace file listing~4k~4k †~4k †
message history~1k~4k †~8k †
new message~2k~3k~2k
Turn total~16k~20k~23k
Cumulative in the minute16k36k → 42959k
† re-sent, byte-identical. Sizes are illustrative of a ~15–25k-token turn; the limit (30k input tokens/min) is Tier 1.