Skip to main content
Black tier only. One shared budget covers AI Auto-Moderation and AI Server Control — there is no separate allowance per feature.

​
Two pools

Your balance is the sum of two pools, spent in this order:

Monthly grant

10,000,000 tokens, refilled on a 30-day cycle. Unused tokens do not roll over.

Top-up

Tokens you have purchased. They do not expire and carry over when the subscription renews, but they are lost if the subscription ends or the server drops below Black.
The grant is always consumed first, so a top-up is a genuine reserve rather than something that quietly burns down each month.
“Resets in” counts down to the next grant refill, which comes every 30 days whatever your billing period — a yearly subscription refills every 30 days too. It shows ”—” when no cycle is running: the grant is full, and the next 30-day cycle starts with your next AI request.

​
Overview

Days roll over at UTC midnight, not your server’s timezone.
The monthly figures are cached for up to five minutes. Immediately after heavy activity the daily tiles update before the monthly ones — they are not disagreeing, just refreshing at different rates.
The Token Usage chart plots input and output as a stacked area over 7, 14 or 30 days. Days with no activity show as zero rather than gaps. Top Token Consumers lists the ten members whose activity cost the most today, with their request count and average cost per request. It only appears once there is something to show.
A member sitting at the top of that list day after day is usually posting a lot of images or videos. Auto-Mod scans up to four images per message and makes up to four calls per video, and both cost far more than text.

​
Activity

Every AI request, newest first, 20 per page. AI Server Control rows expand to show each action it ran — a tick for success, a shield for a refused action, a cross for an error — and how long each took.
Request history is kept for 30 days. The chart and export reach back up to 90 days, so older spend still appears in the totals after the individual rows have gone.
A Server Control run that fails, is blocked or times out after the model has produced tokens is billed and listed — a timeout or error shows as Failed. A run that could not reach the AI at all costs nothing, and is listed as Failed with no tokens.Each row is written right after its charge. If that write ever fails, the spend still counts in the totals without a matching row.

​
Export

Export CSV downloads the last 30 days — one row per day with input, output and total tokens, request count and action count. The chart’s period selector does not change the export window.

​
Topping up

$1 per 500,000 tokens, in blocks of 1, 5, 10, 25, 50 or 100 units.
1

Choose an amount

The dialog shows the token total and price as you select.
2

Complete checkout

You are sent to Paddle to pay.
3

Tokens are credited

The purchase is added to your top-up pool once payment confirms.
The Top Up Tokens button only appears if the server has an active Black subscription billed through Paddle. Servers on a granted or complimentary Black tier have no payment method on file and cannot top up.
The balance does not update the instant you pay. Tokens are credited once the payment provider confirms the payment, so refresh the page after a minute. If AI features had paused, they resume as soon as the tokens are credited.
Every purchase is credited exactly once, even if the payment provider sends the confirmation more than once. Refunds debit the tokens back.

​
When the budget runs out

The two features stop at different points, because a Server Control run needs far more headroom than a single message scan.
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.
Notices are posted to your Auto-Mod alert channel, falling back to the log channel.

​
Keeping spend down

The single biggest saving. A message with text and four unseen images usually costs five AI calls instead of one — one for the text and one per image, plus one more for any image the first model refuses to judge, since a second model then checks it. Reading the text in an image costs nothing, and images the server has already scanned are usually answered from cache at no cost; a message with images but no text runs no text call. Videos ride the same switch: each new video costs up to four calls — up to two for its frames, one to transcribe its speech and one to check what was said and shown on screen, which is free when there is no speech.
Long-standing clean members get sampled rather than scanned every time — while invites, mass mentions and links are still always checked.
A bot-command channel, or a memes channel among trusted regulars, rarely needs paid scanning.

​
Troubleshooting

Reload the page. If it persists, the subscription record for this specific server may not be resolving — check the server appears under your subscription on the Premium page.
Give it a minute and reload. If AI had paused, it resumes as soon as the tokens are credited.
That is a sign-in failure on the download. Reload the dashboard so your session refreshes, then try again.
Check the list’s filters first — a status, user or search filter narrows the rows without changing the totals above them. Otherwise it is the retention gap: rows are kept for 30 days while the chart and totals reach back 90, so older spend has no row left to explain it.Both features write a row per billed call, so Auto-Mod’s scans of messages, images and videos are in the list too; image- and video-heavy activity is what usually makes the number grow fastest.
No — there is one shared pool. If Auto-Mod is consuming it, tighten its exemptions or Trust Tiers rather than looking for a per-feature cap.

AI Auto-Moderation

Where most token spend comes from.

AI and tokens

How the Black-tier AI allowance works.