Quick answer
A chat turn is one model call. An agent run is a loop of model calls plus tool use, so it can consume many times the credits. Use chat for questions and small edits, and agents for multi-step tasks where the loop saves you more time than it costs.
The cost difference
| Mode | Model calls | Typical use | Credit impact |
|---|---|---|---|
| Inline completion | One | Finish a line | Included or minimal |
| Chat | One per message | Explain, draft, review | Low |
| Agent | One per step | Multi-file task | High |
| Agent with retries | Many | Debugging loop | Very high |
Choose the cheaper mode
- Ask a question in chat before starting an agent.
- Use an agent when the task needs file edits and test runs.
- Stop the agent when the same error repeats.
- Start a new agent session per task to avoid replaying context.
- Use the default model for mechanical edits.
Measure the split
Organization metrics separate chat turns from agent activity, which is the first place to look when credits rise without more people. On a personal account, compare the balance before and after a session.
UsageAtlas keeps the reported Copilot balance over time so the jump is visible.
Chat and agent consume very differently
| Mode | Model calls | Allowance impact |
|---|---|---|
| Inline completion | One short call | Minimal |
| Chat question | One call plus context | Moderate |
| Agent task | Many calls plus tools | High |
| Agent loop on failure | Repeated calls | Highest |
The same task can be cheap in chat and expensive in agent mode, because the agent reads files, runs commands, and revises. That is a fair trade when it finishes the task; it is pure cost when it loops.
Set a stop condition before starting an agent run, and see limits by plan for what each tier allows.
Questions, answered
Does an agent run count as one premium request?
No. An agent run can make many model calls, and each one can consume credits.
Is chat cheaper than inline completion?
Inline completion is usually included on the default model. Chat and agent calls are the metered part.
Can I limit agent steps?
Some clients allow a step or time limit. Where they do not, stop the run manually when it loops.