Start with messages, not registered accounts
A thousand registered accounts do not tell you how many model calls you will pay for. Some people may never open support; others may send several messages in one conversation. Estimate the number of people using the feature each month, then the messages each person sends.
In the worked example above, every listed monthly user sends the same average number of messages. If your app has a larger audience but only a fraction uses the assistant, enter the people who use it, or adjust the average across your whole audience. Do not apply that fraction twice.
A message means one user action that triggers the configured number of model calls. It is not a resolved ticket or an entire conversation. The example makes one call per message. A workflow that classifies a request, drafts a response and checks it may need several calls; account for those before relying on the estimate.
What is inside each model call?
Input includes the instructions, the user’s message and any context you send with it. That context might contain previous turns or passages from a knowledge base. Output is the generated response. The table above makes both quantities explicit so you can replace them with a representative sample from your app.
The example uses a fixed average per call. It does not simulate a conversation whose history grows at every turn. If long sessions are common, measure or estimate that larger average. Output that looks short to the user may also require billable reasoning; check your chosen provider’s accounting and include it in the workload where applicable.
For each workload, the engine prices input and output separately using the selected model’s applicable rate band. The calculation trace shows the actual formula. It does not average a small prompt with a large one before choosing the context band.
Read the three usage levels
The usage table keeps the per-person assumptions fixed and changes the monthly audience. Each row is a new calculation against the same catalog version. This shows the effect of volume without quietly changing the task.
The selected option is the lowest-cost complete compatible stack with fresh prices in the supported catalog. Low cost does not prove that a model will answer support questions correctly. Test candidate models on representative tickets and your escalation rules. In the calculator, Quality fit is a separate evidence-based view; it is not a guarantee for your own support data.
To inspect provider coverage, start with OpenAI, Anthropic or Google. Their pages show the specific offers we can price, rather than every product those companies sell.
What this estimate leaves out
This support example prices the model calls. It does not include your application hosting, storage, vector database, embeddings, human support staff or engineering time. Supplying retrieved text as model input does not account for the cost of retrieving or indexing that text.
Retries and extra calls are not added automatically. Caching, batch discounts, negotiated rates and taxes are outside this example. Recurring free allowances are switched off. Check the methodology before treating the estimate as a project budget.
If your assistant searches the live web or reads external pages, use the research-agent guide to understand those separate API charges. You can combine Chat and Research in the calculator, but make sure you do not count the same model call twice.
Turn the example into your budget
Open a row from the table and change the audience and activity first. Then inspect the token assumptions and provider constraints. The link preserves the example’s inputs and data versions so you can see what changed. A saved price may age; review its date before choosing a service.
Use the estimate to narrow the options and plan a small real workload test. Compare the resulting usage with your assumptions, then update the budget. That is a stronger basis for a launch decision than assuming every conversation will behave like the demo.