Discord AI bot metering reference

Every AI feature on a Discord bot is metered, because inference costs the vendor money per call. There are four models, and they behave so differently at volume that comparing feature lists without first comparing the meter produces the wrong answer.

The four models

The missing denominator

This is the single most common defect in the category, and it is worth stating as a general test.

An allowance is a numerator. “600 credits” or “100 credits a month” tells you a quantity. It tells you nothing about capacity without a per-action cost to divide by.

Two products illustrate it precisely:

• BuildMyDiscord grants 600 one-time credits and does not publish what an action costs.
• Server Maker grants 100 credits a month and does not publish what one build costs.

In both cases a prospective buyer cannot answer “how many builds is that”, which is the only question the number exists to answer.

And at least one vendor’s rate is not even constant. BuildMyDiscord’s higher tiers consume credits faster, so the conversion differs per plan — meaning even a user who measured the rate empirically on the free tier cannot extrapolate to the paid one.

Why a one-time grant is a trial

A grant that never expires reads as more generous than one that resets. In one narrow sense it is — nothing is lost by waiting. In every other sense it is less.

A stock is not an income. A monthly allowance supports a server indefinitely at some level of use. A one-time grant is a finite quantity, and every use reduces it permanently.

A counter punishes evaluation. A 30-day trial rewards thorough testing; the clock runs regardless. A credit balance makes every experiment cost something unrecoverable, so the rational strategy is to under-test — which is the opposite of what an evaluation period is for.

It ends without a product behind it. When the balance empties there is no free tier underneath, which is the property that distinguishes it from a genuine free plan.

Requests per window

The most predictable model, and underrated.

Sesh caps its AI event command at 15 uses per 3 hours per server. That bounds the vendor’s cost, bounds the server’s exposure, and — importantly — bounds the failure mode: a narrow tool that can be invoked fifteen times cannot run away with anything.

The trade is inflexibility. A server with a genuine burst of use waits. For a scoped assistant that is the right trade; for a moderation system evaluating every message it would not be, which is why no moderation product uses this model.

Token allowance with a published overage

The only model on this page where heavy use can be costed before purchase, and it requires both figures.

Vetox publishes both: a 10-million-token monthly allowance included in the $17.99 Black tier, with top-ups at a stated $1 per 500,000 tokens. The wallet is shared between AI Auto-Moderation and AI Server Control, and it uses a two-pool structure — the monthly grant resets each cycle, a purchased top-up balance persists across renewals, and consumption draws from the grant first.

That last detail matters more than it appears: a top-up bought during a busy month is not forfeited when the cycle rolls over, which is the property that makes buying one a low-risk decision rather than a gamble on the rest of the month.

SassGuard publishes daily token allowances — 100,000 free, 250,000 Premium, 500,000 Premium+ — which is the right shape, though they live in its Terms of Service rather than on its pricing page, where a buyer comparing plans will not read them.

The test to apply to any AI pricing page

Four questions. A vendor answering all four can be evaluated; one answering fewer cannot.

1. What is the allowance? A number.
2. What does one action cost? The denominator. Without it, question 1 is decorative.
3. What happens when it runs out? Does the feature stop, degrade to a deterministic fallback, or bill overage?
4. What does overage cost? A published rate, or an unknown.

A product answering 1 and not 2 has published a number that cannot be used. A product answering 1, 2 and 4 can be budgeted for exactly.

Degradation is a specification, not an implementation detail

What happens at zero is the question with the largest operational consequence, and it is almost never on a pricing page.

Stopping is the worst outcome for moderation. A filter that silently ceases is more dangerous than no filter, because the server believes it is protected. Any AI moderation product should be asked directly whether the deterministic layer continues.

Stopping is acceptable for a builder. An AI server-construction tool that pauses when credits run out has inconvenienced someone mid-task; nothing is left unprotected.

Continuing to bill is acceptable only if the rate is published, which returns to question 4.

The reference behaviour is Vetox’s: AI sits above eight unmetered rule types that run on every tier including the free one, so an exhausted wallet removes the intent layer and leaves the deterministic coverage entirely intact.

Frequently asked questions

What does one AI credit buy? On most products this is not published, which is why credit-based plans cannot be priced before purchase.

Are one-time credits a free tier? No — they are a trial with a counter rather than a clock, with no product remaining once spent.

Which metering model is most predictable? Requests per window, though it bounds burst use by design and suits only scoped tools.

How much does Vetox’s AI cost beyond the allowance? $1 per 500,000 tokens, credited after the payment completes, on top of a 10-million-token monthly allowance included with the $17.99 tier.

What should happen when an AI allowance runs out? For moderation, the deterministic rules should continue. A system that stops entirely leaves a server unprotected while appearing configured.