Keeping costs down¶
What spends extra requests¶
Some settings make one more model request each time. Each says so where you switch it on.
| What | When it spends |
|---|---|
| Note what the work found, by itself (Settings › This project › Memory and web) | One request after every message. |
| Have a model look over each sub-agent's report (Settings › This project › Agents) | One request after every sub-agent's report. |
| Orchestrator model: automatic (the chat's model picker) | One request for each message you send. |
| The Orchestrator answers what it safely can (Settings › This project › Asking and away › When you're away), on a CLI | One request of your plan for each answer. |
| Try a real run (Models page, a CLI) | One short run, on your plan. |
| Read now for Claude Code (a provider's card) | One short request. |
| Search by meaning (Memory page) | One embeddings request for each search. |
| Learning a knowledge topic (Knowledge page) | Model requests, up to the limit you set. |
| Candidates in a recipe | Each candidate is its own agent. |
| Agents on a CLI working at the same time | They spend your subscription at the same time. |
Ways to keep a job cheap¶
- Give models tiers, and let roles ask for a tier, so easy jobs go to a fast model. See Tiers.
- Turn off Agents may use for an expensive model you want to pick only by hand.
- Set a Monthly budget and a token limit per message (Settings › This project › Limits and money), and dollar limits on API keys (a provider's Limits tab).
- Keep Agents at the same time and Sub-agents per request low (Settings › This project › Limits and money).
- Give an MCP server only to the roles that need it, and switch off tools you do not use: every tool's description is sent at every model step.
- Use a model server on your own machine for work that does not need a large model: it spends no money.
- On GitHub Copilot CLI, give the provider a model. Copilot's
autopicks one for each request; with no model and Medium thinking, Copilot's default spent 7.12 AI credits for the same short answer thatautoanswered for 0.18 (measured once, 7 Nov 2026, Copilot CLI 1.0.105). See GitHub Copilot CLI.
See Usage and budgets.