How to run AI generation jobs with cost awareness

Find the provider with a query for the generation capability, for example the fal.ai plugin. Fetch the exact interface from the directory before you send. The fal.ai plugin declares image-generation, speech-audio-generation, speech-to-speech, music-generation, embedding-generation and model-catalog. No video generation interface exists, so say so instead of trying a workaround.

Choose a model first

Send query-models on model-catalog with the capabilityInterface you need. Each entry gives the model ID to pass, its supported aspect ratios and resolutions, and prompting instructions. An unknown model is rejected, so never guess an ID. generate-image takes model and prompt, and optionally aspectRatio, resolution and responseFormat. Prefer URL, because BASE64 fails above 10 MiB. Inspect the returned media before you describe it; the URL is short-lived.

Cost and confirmation

Every run is billed to the customer's provider account. The interfaces and the catalog publish no price, balance or budget, so you cannot read or enforce a spend limit. Do not state a price as fact. Say that cost depends on the model and the resolution tier, and that higher tiers cost more.

Before any run, tell the customer the model, what you will generate, how many runs, and that it may cost money. Ask for confirmation. Always ask first for batches, retries, a higher resolution, or several models to compare. When the customer is exploring, start with one run on a suitable lower-cost model.

A run takes seconds to about two minutes. Send it once. Accepted control is not a finished result, and a timeout leaves the outcome unknown because the job may still be billed. Do not resend automatically; check the outcome and tell the customer what happened.

Read authorization rules and async outcomes.