Why agents cost more than chats
An agent makes many model calls per task — plan, act, observe, repeat. The same task a chat answers in one shot can cost an agent several times more, and that is normal. What matters is whether each of those calls earned its cost.
The levers, biggest first
- Smaller model for routine steps: draft and summarise on a cheap model, reserve the big one for hard reasoning.
- Cap everything: token budgets and daily spend caps turn a bug into a stopped run instead of an invoice.
- Shorter context: compact old turns instead of replaying the whole conversation.
- Local models where it fits: Ollama costs $0 per token for tasks that do not need a frontier model.
- Fewer verification loops: tell the agent what “done” means so it checks once, not five times.
- Schedule, do not babysit: a nightly cron job does not sit in an idle chat session.
- Read the numbers: track token spend per task and cut the repeat offenders.
- One fleet budget: with several devices, a single spend alert catches all of them.
What not to cut
Do not cut the audit trail or disable secure mode to save cost — those are the controls that make the spend safe to begin with. A cheap uncontrolled agent is the most expensive kind.
A realistic monthly target
Record your current monthly cost, apply two levers, and re-measure after a week. The goal is cost per finished task going down, not raw spend — a cheaper agent that never finishes anything is not a saving.