# AI memory cost worksheet

Publisher: Gigabrain Labs · Prepared 2026-09-05

Use actual observed usage and verified prices for a buying decision. The example below uses invented rates and volumes solely to demonstrate arithmetic.

## Runnable example

[Download the standard-library Python calculation](https://gigabrain.download/assets/resources/calculate-memory-cost.py) and [its recorded output](https://gigabrain.download/assets/resources/memory-cost-result.json).

Run `python3 calculate-memory-cost.py` with Python 3.11 or later. The script has no network access or dependencies.

For 100 tasks, A uses 12,000 input and 800 output tokens per task; B uses 3,000 input and 800 output. Hypothetical rates are $1 per million input and $4 per million output tokens. A costs $1.52 and B $0.62 for those categories. Adding $2 of extra processing to B produces $2.62, which is $1.10 more than A. These are not current prices, measured product usage or promised savings.

## Your comparison

| Item | Current workflow | Proposed workflow | Evidence / assumption |
|---|---|---|---|
| Period and completed tasks | | | |
| Software acquisition | | | |
| Recurring software / updates | | | |
| Input usage by billed class | | | |
| Output and other model usage | | | |
| Initial indexing | | | |
| Changed-source processing | | | |
| Retrieval / hosting / storage | | | |
| Backup and recovery | | | |
| Retries and unsuccessful tasks | | | |
| Actual human setup and review | | | |
| Correct, usable completed tasks | | | |

## Interpretation

Separate one-time costs from recurring costs. Record the period over which you spread setup work. Avoid treating all elapsed time as paid labor or savings. Compare task quality and effort alongside cost; a small pilot does not justify a precise annual return forecast.