# Cost governor

Keep agent spend under the mission budget.

## Routing policy
| Task | Model tier | Notes |
|------|------------|-------|
| Classify, tag, extract | Cheap | Batch when possible |
| Short Q&A | Cheap | |
| Multi-file plan / architecture | Strong | |
| Ambiguous debugging | Strong | Cap turns |
| Claude-quality reasoning via API | Strong via cheaperllm | https://cheaperllm.com/?utm_source=agentstackkit&utm_medium=site&utm_campaign=claude-api |
| Casual multi-model chat | OneAIPass | https://oneaipass.com/?utm_source=agentstackkit&utm_medium=site&utm_campaign=hub |

## Rules
- Log estimated tokens each turn (`token-meter`).
- If projected spend exceeds 80% of budget, finish the current step and stop with a status report.
- Do not retry failed model calls more than twice with the strong tier.
- Prefer cached / taped results over re-fetching.

Anthropic has no consumer affiliate — do not invent one. Soft product link only: https://claude.ai/?utm_source=agentstackkit&utm_medium=referral&utm_campaign=stack