Each night sends log excerpts plus short summarising instructions. It returns a compact report, so input again outweighs output.
Output length varies by task and model. Budget from measured report sizes, not from hope.
Period | Nights | Cost (units) |
|---|---|---|
Per night | 1 | 0.204 |
Per week | 7 | 1.428 |
Per 30 nights | 30 | 6.12 |
One night carries 60000 input tokens and 2000 output tokens. The input part is 0.18 units and the output part is 0.024 units.
EXAMPLE prices (not real prices): 3 units per 1M input tokens, 12 units per 1M output tokens
one night: 60000 input tokens, 2000 output tokens
input part = 60000 / 1000000 x 3 = 0.18 units
output part = 2000 / 1000000 x 12 = 0.024 units
per night = 0.18 + 0.024 = 0.204 units
input share = 0.18 / 0.204 = 88.24%
output share = 0.024 / 0.204 = 11.76%
88.24 + 11.76 = 100.00Thirty identical nights cost 30 times one night: 6.12 units per month. That product is the monthly budget line.
Halving the excerpt window to 30000 input tokens gives 0.09 plus 0.024, which is 0.114 units a night and 3.42 a month. The saving is 2.70 units a month against the same reports.
Seven nights cost 1.428 units. Thirty nights cost 6.12 units.
The input part halves to 0.09 units. The night costs 0.09 + 0.024 = 0.114 units.