A team needs 100,000 usable answers a month. Should it pay per API call or rent a physical server? Give both options the same Georgian questions, acceptance criteria and response deadline. Compare the cost per 1,000 answers that pass.
Fill in your monthly estimate
| Input | Your figure |
|---|---|
| Accepted answers / billed attempts | ___ / ___ |
| Input / output tokens per attempt | ___ / ___ |
| Price per million input / output tokens, one currency | ___ / ___ |
| Rental / extra backup / licences per month | ___ / ___ / ___ |
| Administrative hours × hourly cost | ___ × ___ |
| Peak requests per minute / acceptable latency | ___ / ___ seconds |
A token is a unit used to meter text. Get actual input and output counts from your Georgian test requests. Use monthly totals if lengths vary. Account for caching, tools and billed retries when applicable. Example of an API pricing schedule.
Worked example, not a commercial quote
All figures are invented, in GEL and inclusive of tax. There is no multi-month prepayment. API usage is billed monthly in arrears, while the physical rental is prepaid for one month. Equal quality-review costs are omitted on both sides.
API. There are 120,000 paid attempts for 100,000 accepted answers, including 20,000 retries. At 2,000 input and 500 output tokens per attempt, the month uses 240 million input and 60 million output tokens. Hypothetical metered rates per million, billed monthly without prepayment and inclusive of tax, are 2 GEL in and 12 GEL out. The input costs 240 × 2 = 480 GEL. Output costs 60 × 12 = 720 GEL. Total: 1,200 GEL, or 12 GEL per 1,000 accepted answers.
Physical server. One month’s prepaid rental is 720 GEL including tax. Add 60 GEL for that month’s extra backup, zero for extra licences and six hours of administration at 30 GEL, or 180 GEL. Power and the agreed network are included in this example’s rental. Total: 960 GEL, or 9.60 GEL per 1,000 at the same volume.
Find the intersection
With the same response lengths and retries, the API costs 0.012 GEL per accepted answer. The server stays at 960 GEL within tested capacity. 960 ÷ 0.012 = 80,000 accepted answers a month. Both cost 960 GEL there. At 40,000 answers, the API costs 480 GEL and the server still costs 960 GEL.
A lunch-hour peak may need a second machine even while the first sits idle at night. Recalculate the fixed cost in that case. Record peak requests per minute, responses meeting the deadline and editorial rejections. A cheaper total cannot solve a growing queue.
A dedicated CPU server can prepare data, host retrieval or run a model that meets your measured target on CPU. If model computation dominates, request the same test on a GPU configuration. Confirm the hardware, quote and delivery separately. Use the GPU memory guide to size the trial. Send the accepted monthly volume, tokens in both directions, busiest minute and target latency for a configuration proposal.



