Skip to main content

Estimated Savings

PaletteAI Inference Launchpad estimates how much you save by serving inference on the appliance instead of sending the same traffic to a frontier model. The estimate is a comparison, not money returned. This page explains how the appliance calculates it and where to find it.

Estimated savings build on the estimated cost of each request, which depends on the four per-model rates and on how the appliance prices cache reads. For those, refer to Usage Pricing.

How Estimated Savings Are Calculated​

The comparison is against two fixed frontier models, Anthropic Claude Opus 4.8 and OpenAI GPT-5. Client model aliases do not change which models the comparison uses.

For every request served on the appliance, the appliance prices the same token mix at the comparison model's rates, using the same split between fresh and cached tokens that the meter uses.

comparison estimate = fresh input tokens × comparison input rate
                    + cache-read tokens  × comparison cache-read rate
                    + output tokens      × comparison output rate

savings = comparison estimate − estimated cost on the appliance

The comparison cache-read rate is the comparison model's Cache read rate. At the provider list prices, that rate is one tenth of the comparison model's input rate. Filling a cache costs nothing on the appliance, so the comparison has no cache-write term for locally served traffic. The estimated cost on the appliance is the request's metered cost at the hosted model's rates, or $0 for a response cache hit.

Estimated savings are never negative. A locally priced model whose rates exceed the comparison model's reports zero savings, not a loss.

The comparison model's rates are the same provider list prices the appliance uses to price egress, so the savings estimate and the estimated egress cost quote the same numbers. If an operator overrides the comparison model's rates under Settings > Pricing, the savings estimate uses the override.

Example: Estimated Savings​

This example uses the following illustrative rates for the Opus 4.8 comparison. These are example rates, not current list prices.

  • Input: $5.00 per 1 million tokens.
  • Cache read: $0.50 per 1 million tokens.
  • Cache write: $6.25 per 1 million tokens.

A locally served request sends a prompt of which the engine answers 400,000 tokens from cache and computes 500,000 fresh, with no output. To isolate the comparison, the example sets the request's estimated cost on the appliance to $0.

StepCalculationAmount
Fresh input at the input rate500,000 × $5.00/M$2.50
Cache reads at the cache-read rate400,000 × $0.50/M$0.20
Comparison estimate$2.50 + $0.20$2.70
Estimated savings$2.70 − $0$2.70

The comparison has no cache-write term. The appliance never charges to fill its own cache, so the comparison does not charge for it either, and the example's cache-write rate does not apply.

Where to Find Your Savings​

To view your savings, select Overview from the left main menu. The Estimated savings card reports two amounts, saved vs Opus 4.8 and saved vs GPT-5, for the period selected on the page. A chart beside them plots both amounts across that period. The card does not appear when the appliance served no requests in the period. The Edit Model Pricing link on the card opens Settings > Pricing.

The By Client export on the Usage page also carries both amounts in its header. For the export layout, refer to Usage Metrics Reference.

Resources​