Open almost any app you use weekly and you will find an image generation button somewhere. Your note-taking app makes cover art. Your project tool generates avatars. The presentation software you tolerate builds slide backgrounds on demand. Two years ago this would have been a headline feature; today it is a checkbox, and the interesting question is no longer whether an app has it but how a small team could afford to ship it.
The Part Nobody Explains
The assumption most people carry is that AI image generation is expensive — that behind every button press sits a rack of GPUs someone is paying dearly for. That was briefly true. It stopped being true when the model labs opened metered access, and the economics quietly inverted.
A single generated image today costs somewhere between a fraction of a cent and a few tens of cents, depending on resolution and quality settings. Not per month. Per image. A hobby project generating a few hundred images a month is looking at a bill smaller than a streaming subscription.
That is the shift that let indie developers ship features that used to require a funded team. The capability stopped being a moat.
How It Actually Gets Wired In
The mechanics are less dramatic than the output suggests. An app sends a text description to a hosted model over an ordinary HTTP request and gets an image back. No local hardware, no model weights, no machine learning expertise on staff. If you can call a weather API, you can call an image model.
Where developers spend real thought is on which model to call. The landscape moves fast enough that hard-wiring one vendor is a bad bet — the best model for photorealistic scenes is rarely the best for flat illustration or for text inside an image. This is why the pattern that has settled in for most small teams is an aggregation layer: one account and one endpoint fronting many models. Published rates for the GPT Image 2 API sit alongside competing image models on those platforms, so swapping the engine behind a feature becomes a config change rather than a rewrite.
The Cost That Actually Bites
Here is the number that decides your bill, and it appears on nobody’s pricing page: how many times you generate before you accept a result.
Nobody keeps the first image. Real usage runs three to eight attempts per keeper, which means your true cost per usable image is several times the headline figure. The teams who keep bills sane do one unglamorous thing — they generate cheap low-resolution drafts to pick a direction, then regenerate only the winner at full quality. It halves the spend and nobody can tell from the output.
What This Means If You Are Building Something
The barrier to adding generated imagery to a side project is now a weekend and a small budget. The barrier to doing it well is unchanged: knowing what the image is for, writing a prompt that describes it, and having the judgement to reject the seven attempts that missed.
The technology got cheap. Taste did not. That gap is where the apps worth using still separate themselves from the ones that just shipped a button.
