Budget for AI search and tool-call costs
On this page
Budget model usage and separately charged grounding or tool usage as distinct lines. Count how many billable operations one user action triggers, then apply the current rate for each exact provider SKU or contract. Do not apply one token rate to the whole workflow.
Why this is worth a look
The sources show that an AI workflow can have distinct charge or usage dimensions. Vertex AI lists input and output token prices, explains per-turn context-window billing, and states that Google Search grounding is billed separately from Gemini model usage when a prompt successfully returns sources. AgentCore documents browser and code-interpreter session duration plus vCPU-hour and memory GB-hour usage telemetry. Azure's Web Search tool uses Grounding with Bing and has additional costs and terms. Model the operations you actually use, and verify each selected service's billing unit before assigning a rate.
Run this check
CHECKLISTA read-only worksheet for one workload and billing period. Fill it with observed or load-test units and the current rate for each exact provider SKU. No account access, permissions, resource changes, or external columns are required.
Workload and billing period: ____________________
Scope: account/project/subscription and region(s): ____________________
Expected user actions: __________
Peak and average request rate: __________
Model operations per user action
initial model calls: __________
model calls after tool results: __________
retry attempts: __________
evaluation calls: __________
input tokens by model, region, and context tier: __________
output tokens by model and processing mode: __________
context-cache storage or other cache units, where separately billed: __________
Tool and grounding operations per user action
search or grounding requests, using the selected service's documented unit: __________
retrieval or data-processing units, if separately billed for this service: __________
browser sessions and duration: __________
code-interpreter sessions and duration: __________
other tool calls by documented billable unit: __________
Rate card or contract captured on: __________
[ ] exact provider, model, region, deployment, context tier, and processing mode
[ ] documented input and output rates
[ ] documented search or grounding rate and charging condition, if applicable
[ ] documented tool session, CPU, memory, storage, or data-processing rate, if applicable
[ ] provisioned-capacity or commitment terms, if applicable
[ ] cache rate categories only where the selected service documents them
Budget formula for each documented line:
period cost = user actions x operations per action x units per operation x current unit rate
Add separate lines for fixed provisioned charges.
Add monitoring or log-ingestion lines only if separately billed in this environment.
Compare the estimate with the provider billing statement or invoice for the same period.How to confirm it
- 01
Map the billable workflow
For each user action, list model calls, search or grounding requests, retrieval or other service operations if separately billed, browser or code-interpreter sessions, model calls that consume tool results, retries, and evaluations. Include branches and loops, not only the successful path.
- 02
Measure units in scope
Use application traces or the selected service's usage records to count the documented units for the same account or project, services, regions, and billing period. AgentCore session-level logs can expose duration, browser vCPU-hours and memory GB-hours, and code-interpreter vCPU-hours and memory GB-hours. Use distributions when the workload varies, rather than one average alone.
- 03
Match rates to documented meters
Record the exact provider, model, region, context tier, deployment or processing mode, grounding feature, and tool. Apply separate model and grounding lines where the provider documents them. Treat cache categories, retrieval, data processing, allowances, and other meters as applicable only when the selected service documents them.
- 04
Model normal and high-volume cases
Calculate a baseline from observed typical usage and a high-volume case from higher observed operation, retry, token, and session counts. Keep provisioned capacity or commitment terms separate from on-demand usage. Use consistent units for every line.
- 05
Reconcile against authoritative billing
Compare estimated units and rates with the provider billing statement or invoice for the same scope and period. If a billing export is also used, verify that it covers the same account, services, adjustments, and period. Investigate differences before changing the model.
Before making changes
Tool charging differs by provider and feature. Azure's Web Search tool uses Grounding with Bing, which has additional costs and terms, and customer data flows outside the Azure compliance boundary. Google Search grounding is billed when a prompt successfully returns sources, with Gemini model usage charged separately. AgentCore telemetry is for monitoring and can differ from metered billing. These assumptions require the selected service's documented units, scope, and billing period, and do not assume universal cache, retrieval, allowance, or contract meters.