Costs and usage
Costs depend on the models, cloud resources, and services you actually use. Start with a small job and inspect usage before scaling up.
On this page
Possible cost categories
| Category | What it covers |
|---|---|
| Model usage | Processing inputs and generating outputs |
| Agent rates | Pricing configured for third-party assistants |
| Sandbox and storage | Cloud working environments and retained data |
| Application resources | Cloud functions, databases, and related usage |
| External orders | Specific products or paid services |
Swipe sideways to see more of a wide table.
Use the current console, confirmation screens, and billing records for rates and settlement terms. This guide does not promise a fixed price or free allowance.
What is a Token
A Token is a model's unit of input or output measurement, not exactly a word or character. Conversation history, tool results, and generated responses can contribute to usage beyond your latest message.
A short “continue” can cost different amounts in different contexts. Models and task steps also affect usage.
Check your bill
Open Billing to inspect balance, top-ups, gifts, and consumption. Use available date, category, or member views and verify the current organization.
Compare usage before and after a larger job instead of estimating only from reply length. Insufficient balance, resource quotas, and an unavailable model are different problems.
Sleep and retained resources
Idle work sandboxes can sleep to reduce unused compute. Retained files and snapshots can still incur storage costs. No active conversation does not necessarily mean no resource charges.
Shallow sleep pauses an existing environment; deep sleep involves saving and recreating it. Do not assume every process or connection survives unchanged.
Archiving destroys the session sandbox but keeps its drive files. Applications, databases, and schedules are separate resources to inspect and manage.
Keep usage under control
Approve outlines and rules before large batches. Provide relevant materials, choose models based on results, and avoid duplicate delegation. Test recurring tasks and pause them when no longer needed.
Regularly review unused environments and application resources. After adding balance or resolving a quota, inspect the task state rather than assuming every stopped operation will rerun automatically.
Before restarting work that sends messages, takes payments, or writes data, check what already happened to avoid duplicates. Record the session ID, time, and error when seeking help.