Possibly related to #39837, the AI credits shown in the Token Usage output do not take into account the cached tokens. Below is a real example.
On turn 1, it correctly computes the cost of the turn, which uses 0 cached tokens obviously, as 6.57 AI Credits.
On turn 2, it computes the delta AI credits as per follows:
27369 Input tokens ($2.5/million): 6.84 AIC
264 Output tokens ($15/million): 0.396 AIC
22528 cached read tokens ($0.25/million): 0.563 AIC
If we add the three together we get the 7.8 AIC shown in the delta in the screenshot, however a majority of those inputs tokens were cached and so were double counted (both as fresh input tokens and again as cached tokens). I believe the correct calculation for turn 2 delta should be, the new inputs tokens, plus cached tokens, plus output tokens. In this case, 4640 new input tokens (27369 turn 2 input tokens - 22528 cached tokens) which is 1.21 AIC, plus cached read tokens which is 0.563 AIC, plus output tokens which is 0.396 AIC for a total delta of 2.1 AIC, not the 7.8 shown.
This double counting compounds down the way on every turn to give a much higher AI Credit spend than was actually incurred.
Copilot v1.0.65
AWF version v0.27.11
CLI version v0.81.6
Possibly related to #39837, the AI credits shown in the Token Usage output do not take into account the cached tokens. Below is a real example.
On turn 1, it correctly computes the cost of the turn, which uses 0 cached tokens obviously, as 6.57 AI Credits.
On turn 2, it computes the delta AI credits as per follows:
27369 Input tokens ($2.5/million): 6.84 AIC
264 Output tokens ($15/million): 0.396 AIC
22528 cached read tokens ($0.25/million): 0.563 AIC
If we add the three together we get the 7.8 AIC shown in the delta in the screenshot, however a majority of those inputs tokens were cached and so were double counted (both as fresh input tokens and again as cached tokens). I believe the correct calculation for turn 2 delta should be, the new inputs tokens, plus cached tokens, plus output tokens. In this case, 4640 new input tokens (27369 turn 2 input tokens - 22528 cached tokens) which is 1.21 AIC, plus cached read tokens which is 0.563 AIC, plus output tokens which is 0.396 AIC for a total delta of 2.1 AIC, not the 7.8 shown.
This double counting compounds down the way on every turn to give a much higher AI Credit spend than was actually incurred.
Copilot v1.0.65
AWF version v0.27.11
CLI version v0.81.6