Figure 1 — Input token cost per API call in a 5-tool-call turn
Without caching
1
System prompt + history 87k tokens
1.0× rate — full price
2
System prompt + history 87k tokens
1.0× rate — full price
3
System prompt + history 87k tokens
1.0× rate — full price
4
System prompt + history 87k tokens
1.0× rate — full price
5
System prompt + history 87k tokens
1.0× rate — full price
5 × full price every call
$$$
vs
With caching
1
87k tokens — cache write
1.25× rate
new tokens
2
87k tokens — cache read
0.1× rate
new tokens
3
87k tokens — cache read
0.1× rate
new tokens
4
87k tokens — cache read
0.1× rate
new tokens
5
87k tokens — cache read
0.1× rate
new tokens
1 write at 1.25× + 4 reads at 0.1×
$
Full-price input (1.0×)
Cache write (1.25×)
Cache read (0.1×)
Uncached new tokens (each call)