Overview
How might we design pre-task cost estimate so people can trust and act on AI output?
When to use
- Essential for AI coding agents, research agents, and long-running automation where forecasting cost and duration before execution prevents budget overruns and sets honest expectations.
When to skip
- Tiny free-tier completions where estimates would add noise.
- Tasks too unpredictable for any honest range (say so, do not fake precision).
- Fully admin-metered enterprise seats where end users never see cost.
Rules
Point estimates with no range or confidence.
Hiding that tool calls or retries can multiply cost.
Starting the run before the user acknowledges the estimate.
Estimates that ignore cached or already-paid context.
Evidence
| Product | Implementation |
|---|---|
| Cursor | Usage and plan cues around agent and model runs. |
| Claude Code | Usage panels that surface spend during sessions. |
| OpenAI platforms | Token and dollar estimates in playgrounds and billing UIs. |
| Devin-style agents | Task scoping with expected duration and resource hints. |