# Lower your bill

A runbook for people and agents. Always read current prices with `harebox rates`.

## 1. Stop surprises first

```sh
harebox budget new --limit 100 --action stop
```

There is no budget until you set one. Budgets can cover an account, a project, a fleet, a label or an API key. Each sandbox also stops itself an hour after it starts, unless you change its auto-stop.

## 2. See where the money goes

```sh
harebox usage --all --group-by tier
harebox copilot forecast
harebox copilot
```

Each Co-pilot recommendation has evidence, a saving range and a risk. Nothing applies unless you ask, with `harebox copilot apply ID`, and changes that could disturb running work are left for you to make yourself.

## 3. Work through these, in order

| Check | How to tell | Action |
|---|---|---|
| VMs that sit idle while billed | CPU stays low between commands | Stop them when the work ends, or lower `--ttl`. Sleep keeps memory but is billed like running, so it doesn't save money |
| Auto-stop longer than the work | Sandboxes stop long after their last command | Lower `--ttl`, or stop them when the job ends |
| Paying for extra starts | `harebox limits` shows `next_start_usd` above 0 | Batch starts, use `placement: "queue"`, or move to a larger plan |
| Oversized | CPU and RAM stay low | `harebox resize ID --size small --allow-restart` |
| Forgotten stopped disks | Old `stopped` sandboxes | `harebox list --state stopped`, then delete with care |
| AI spend | `harebox ai usage` | `--budget` on each prompt |

## 4. Don't over-optimize

**Spot and Reserved don't save money.** Both cost exactly what Shared costs. Spot can be preempted with two minutes' notice, and a reservation bills for its whole term.

**Don't delete to save money without asking.** Stopped disks are free for 14 days, and then move to cheap cold storage.