# Limits

| Limit | Trial | Build $20 | Team $100 | Scale $500 | Fleet $2,000 |
|---|---:|---:|---:|---:|---:|
| Running at once | 2 | 100 | 300 | 1,000 | 2,000 |
| Starts per minute | 5 | 12 | 30 | 65 | 90 |
| Starts per hour | 25 | 60 | 210 | 420 | 600 |
| Starts per day | 75 | 200 | 840 | 1,680 | 2,400 |
| Free start queue | 10 min | 1 h | 6 h | 24 h | 24 h |
| Saved templates per seat | 0 | 10 | 50 | 250 | 1,000 |
| Sizes | small, default | small, default, large | all | all | all |

A start beyond the plan's included number costs $0.03 up to 1.5 times the included number, $0.05 up to 2 times and $0.07 up to 6 times, taking the highest of the minute, hour and day counts, paid from the wallet. Beyond 6 times (3 times in a minute) starts are refused. Send `placement: "queue"` to wait in the free queue instead. The trial can't buy extra starts.

| Limit, on all plans | Value |
|---|---|
| Waiting command | 60 s by default, 3,600 s at most |
| Streaming or detached command | No time limit |
| Output returned per stream | 256 MiB |
| Environment variables | 500 |
| Webhook endpoints | 100 |
| Public ports per sandbox | 100 |
| Sandboxes in one create call | 10,000 |
| Prompt attachments | 32 files, 25 MB each, 64 MB in all |
| API calls per key | 20,000 per minute, or less if the key sets a lower rate |

Read yours with `harebox limits`. Every response also carries `RateLimit` and `RateLimit-Policy` headers.

## Disk and network

Each sandbox's disk and network are rate-limited per vCPU, as abuse limits: 30 MB/s and 1,400 operations a second of disk, and 20 Mbit/s of network. Bursts of up to 10 s run unthrottled, and each machine also gets a one-time 4 GB of disk and 2 GB of network at boot. A `default` size (4 vCPU) gets 120 MB/s, 5,600 operations a second and 80 Mbit/s.

| Size | vCPU | Disk, MB/s | Disk operations a second | Network, Mbit/s |
|---|---:|---:|---:|---:|
| `small` | 2 | 60 | 2,800 | 40 |
| `default` | 4 | 120 | 5,600 | 80 |
| `large` | 8 | 240 | 11,200 | 160 |
| `xlarge` | 16 | 480 | 22,400 | 320 |