Start by separating managed API charges from infrastructure you operate yourself. A managed model API price ordinarily includes the provider's serving infrastructure; adding a second estimate of that same infrastructure would double-count it. Your own orchestration, storage, or tool services may still be billed separately.
Calculate each billing category separately
Use mutually exclusive token counts and the published rate for each category. When rates are quoted per million tokens:
`text Token cost = ( uncached input tokens * input rate + output tokens * output rate + cache-read tokens * cache-read rate + cache-write tokens * cache-write rate ) / 1,000,000
Request cost = token cost + other applicable request charges `
Do not count cache reads or writes again as uncached input. Some providers charge cache creation at the ordinary input rate; others have a separate write rate. Use the provider's billing definitions, model, region, and service tier. Anthropic, for example, publishes distinct input, output, cache-write, and cache-hit prices in its pricing documentation.
Separate request cost from task cost
A completed task can require multiple requests, retries, tool calls, and human review. Sum the billed requests and separately billed tools for the task. Then compare successful outcomes and quality, not just the number of tokens removed. For a method that chooses models on this basis, see cost per successful task.
If you self-host a model, measure the compute, storage, and network costs of serving it. Allocate shared capacity consistently; do not automatically add a managed API token price to self-hosted inference.
Choose an optimization based on the workload
Model selection, output length, context construction, cache reuse, batching, retry behavior, infrastructure utilization, and purchase terms can all affect cost. A smaller prompt can cost more if it breaks cache reuse or causes extra attempts. TokenShift analyzes activity locally where the agent runs and reports derived metadata and analysis results to its control plane.
Frequently asked questions
Can I multiply all tokens by one rate?
Only if the actual billing contract uses a single rate. Otherwise calculate each input, output, and cache category separately using consistent units.
Which cost lever is largest?
That depends on the workload and billing model. Measure the current cost breakdown and validate changes against successful task outcomes.
Should infrastructure and API costs always be added together?
Add only separately incurred costs. Avoid counting the managed provider's infrastructure twice.
Review date
Pricing categories and this calculation method were reviewed September 17, 2026. Use the current rate card for your exact deployment.