AI and infrastructure
Input and Output Tokens: Why One Blended Rate Can Hide Margin Risk
By the Orvlin team · · 2 min read
Input output token billing: what to decide
Two customers send the same total token volume but generate very different outputs. Their provider costs can differ even if your blended customer charge is identical. The metric is easy to explain but may conceal a cross-subsidy.
Preserve model, direction and relevant provider usage categories in cost records. The customer-facing unit can still be simpler, provided its economics are deliberate. Check current provider documentation rather than assuming all tokens share one rate.
An example with the numbers exposed
For a hypothetical cost model, input costs $1 and output $4 per million tokens. One million input plus one million output costs $5, while two million input costs $2. Both workloads contain two million tokens.
The mistake worth avoiding
A blended rate chosen from one pilot workload can fail as behavior changes. Long answers, tools or reasoning patterns may shift the input-output mix. Monitor cohorts rather than relying on an overall average.
A practical checklist
- Record input and output usage separately.
- Link usage to model and provider rate versions.
- Simulate realistic workload mixes.
- Review margin as behavior changes.
Frequently asked questions
- Must customer invoices show both token directions?
- Not necessarily. You can offer a simpler commercial unit, but retain the underlying detail needed to evaluate costs and explain the charging model.