What a Failed Generation Actually Costs You

There is a question people ask before they buy anything metered, and they usually ask it too late: what happens when it does not work? With an image generator it is a fair question, because generators do fail. A provider times out. A model returns nothing. A prompt gets refused.

The answer at Fellowi is short enough to fit in a sentence, so here it is first: you are charged for pictures that arrive, and for nothing else.

Three things can happen

Press generate and exactly one of these follows.

It works. The picture arrives, the coins are spent, and that is the normal case.

It fails. Something broke between us and our image-generation provider. The coins return to your wallet by themselves.

It is refused. The prompt did not pass our own safety rules and never started. Nothing was charged.

Why a refusal costs nothing at all

This is the part that is worth understanding, because it is a design decision rather than a policy written after the fact. Our hard safety checks run before the money moves, not after it.

A prompt that references a minor, or sexual violence, is stopped at our door. No coin is taken, and the text is never forwarded to our image-generation provider at all. There is no refund to lose, duplicate or miss, because there was never a payment. The alternative design, charging first and refunding after a refusal, is the one that generates support tickets, and we did not build it.

If you want the longer version of which prompts get refused and why, we wrote that up separately in why image generators refuse.

Why a failure refunds itself

When a generation does fail, you do not write to support and you do not wait for a human to look at it. The refund is issued automatically, for the full amount, by the same part of the system that took the payment.

It is also issued exactly once. Each refund carries a key tied to that one generation, so if two different processes both notice the same failure, the second one finds the money already back and does nothing. That sounds like plumbing, and it is, but it is the difference between a promise and a promise that holds when something goes wrong twice at the same time.

The same applies when nothing comes back at all. There is a ceiling on how long a generation may sit waiting; past it, the generation is marked failed and the coins return, rather than leaving you with a spinner and a lighter wallet.

The part we will not dress up

Here is the honest limit. A picture that arrives and is simply not the one you had in your head was a successful generation, and it is charged. Refunds cover technical failure, not disappointment.

We think that is the only workable line, because taste has no test. But it does mean the real skill in keeping your costs down is not hunting refunds, it is writing a prompt that lands sooner. Two posts that help more than anything we could say here: how to write image prompts and keeping a prompt library, which is the habit that quietly halves the retries.

So what is the actual unit of cost

A finished picture, not an attempt. That is the number worth doing arithmetic with, and it is why the arithmetic is simpler here than the pricing pages of most generators suggest.

The current coin costs per generation live on our product page, which reads them from the code that actually charges you, so it cannot drift out of date the way a blog post would. If you want the wider picture of what a month of generating looks like, we worked through it in what AI image generation actually costs, and coins themselves are explained in Fellowi coins explained.

One last practical note, since it is about money in a different sense: what you generate does not stay in your account forever. Download the ones you care about when you get them.

The same rules apply on both sides of the studio, in Fellowi Images and in Fellowi Video. A failed clip refunds exactly like a failed picture.