—vs on-demand
—effective $/1M
—capacity used
Effective price as volume grows
The reservation cost is fixed, so every extra token drives the effective per-million price down — until you hit the capacity ceiling and overflow bills at the on-demand rate.
| Monthly volume | On-demand cost | Reserved effective $/1M | Cheaper |
|---|
Reserve capacity only above the break-even
Provisioned throughput — PTUs on Azure OpenAI, provisioned throughput on AWS Bedrock — swaps a per-token bill for a fixed monthly reservation of dedicated capacity. It buys predictable latency and a flat cost, but it only saves money once your steady token volume clears the break-even: the reservation cost divided by the on-demand price per token. Below that you are paying for idle units; above it the fixed cost spreads over more tokens and your effective price per million keeps falling, right up to the capacity ceiling of units × tokens-per-minute × minutes in the month. Push past the ceiling and the overflow bills at full on-demand rate, quietly eroding the discount. Size the reservation to the volume you are confident you will use every month, forecast growth on the API cost forecast calculator, and compare a spend commitment on the committed spend discount calculator.
Frequently asked questions
What is provisioned throughput and when is it worth it?
Provisioned throughput — Azure OpenAI calls it PTUs, AWS Bedrock calls it provisioned throughput — means renting dedicated model capacity for a fixed monthly fee instead of paying per token. You reserve a number of units, each capable of a set number of tokens per minute, and you pay the same amount whether you use it fully or not. It wins only when your steady token volume is high enough that the equivalent pay-as-you-go bill would exceed the reservation. Below that break-even you are paying for idle capacity; above it every extra token is effectively free until you hit the throughput ceiling.
How do I find the break-even for a PTU reservation?
Divide the total monthly reservation cost by the on-demand price per token. That gives the number of tokens at which pay-as-you-go would cost exactly the same as the reservation. Process fewer tokens than that and on-demand is cheaper; process more and the reservation wins — up to the capacity ceiling, which is the units times tokens-per-minute times the minutes in a month. This calculator shows both the break-even token volume and your capacity utilization so you can see whether you are under-using the reservation or about to run out of headroom.
What happens if I exceed my provisioned capacity?
Requests beyond the reserved tokens-per-minute are either throttled or, on most platforms, spill over to pay-as-you-go pricing for the excess. That is usually fine for short bursts, but if your average demand regularly exceeds capacity you are paying the full on-demand rate on the overflow on top of the fixed reservation, which erases the discount. The calculator flags when your monthly volume is above the reserved capacity so you know to add units or leave the spillover on-demand deliberately rather than by accident.