Effort
Low, medium, or high: how much reasoning a flow asks of its model.
Effort is a separate dial from the model. It controls how much a model reasons, checks its own work, and iterates before returning an answer, independent of which model is doing the reasoning.
The three levels
| Effort | Use for | Example flow |
|---|---|---|
| low | Quick, well-bounded answers with little ambiguity. | Draft customer reply, Explain this codebase |
| medium | Tasks needing judgment but with a clear scope. | Review pull request, Campaign brief |
| high | Multi-step work, or anything where getting it wrong is costly. | Implement a ticket, Quarterly board summary |
Effort is usually the biggest lever on both cost and latency, bigger than the model choice itself. A flow set to "medium" when "low" would have scored identically is the most common optimization Verybear proposes, because nobody ever goes back and checks.
Effort can be set per flow, and changed any time you have evidence it's wrong. See Governance → Budgets and alerts for how savings from effort changes show up.