> ## Documentation Index
> Fetch the complete documentation index at: https://docs.vetox.io/llms.txt
> Use this file to discover all available pages before exploring further.
# AI and tokens
> How the Black tier's 10 million monthly AI tokens work, what consumes them, and how top-ups are credited.
AI requires **Black**. One gate covers all of it: [AI Auto-Moderation](/en/moderation/ai-auto-mod), [AI Server Control](/en/moderation/ai-server-control), the token allowance and AI custom rules.
**Max+ does not include AI.** Custom Bot is an add-on, not a tier. Only Black and Black+ have AI.
## The allowance
**10,000,000 tokens per server, per 30 days.**
A token is roughly a fragment of a word. Both what Vetox reads and what it writes count toward the total.
The allowance refills on a rolling **30-day** cycle — the same on a monthly plan and a yearly one. A yearly Black subscription still receives 10 million every 30 days, not 10 million per year.
Because the cycle is 30 days rather than calendar-monthly, it drifts relative to your billing date. The "resets in" figure on the [AI Usage](/en/moderation/ai-usage) page follows your **billing** date and can therefore be badly wrong on a yearly plan — it may read several hundred days while the grant refills within 30.
Watch the monthly grant figure itself instead.
## Two pools
10,000,000 tokens. **Does not roll over** — unused tokens are lost at the reset.
Purchased tokens. **Never expire** and survive renewals and plan changes.
The grant is always spent first, so a top-up is a genuine reserve rather than something that quietly drains each month.
## Topping up
**\$1 per 500,000 tokens**, in blocks of 1, 5, 10, 25, 50 or 100 units — up to 50 million tokens in one purchase. There is no cap on the balance you can accumulate.
Top-ups require a payment method on file. Servers on a staff-granted or complimentary Black tier cannot top up.
Tokens are credited once payment confirms, exactly once even if the provider sends the confirmation more than once. Refunds debit them back.
## What consumes tokens
| Source | Cost |
| -------------------------- | --------------------------------------- |
| **Auto-Mod text scan** | One call per message |
| **Auto-Mod image scan** | One call per image, up to 4 per message |
| **Auto-Mod username scan** | One call per join or name change |
| **Server Control request** | One or more calls, up to 8 steps |
Image scanning is by far the largest cost. A single message with four images costs five calls instead of one. If your budget is draining, turn off image scanning before anything else.
## When the balance runs low
The two features stop at different points, because a Server Control request needs far more headroom than a single message scan.
| Remaining | Effect |
| ------------- | ----------------------------------------------------- |
| Below 100,000 | Auto-Mod posts a one-time "running low" notice |
| Below 30,000 | **Server Control declines new requests** |
| Below 3,000 | **Auto-Mod stops scanning** and posts a paused notice |
Auto-Mod is fail-closed. If it cannot confirm there is budget it does not scan — your server has no AI moderation until the balance recovers.
## Reducing spend
The single biggest saving available.
Long-standing clean members get sampled instead of scanned every time, while invites, mass mentions and links are always checked.
A bot-command channel rarely needs paid scanning.
A category that is off is never evaluated.
Track spend and top up.
Where most spend comes from.