How to Handle AI Image Generation Timeouts Without Duplicate Charges

When a backend request to an AI image provider times out, developers cannot confirm whether the provider accepted the task or not, creating a risk of duplicate charges and accounting errors. Blindly retrying can trigger a second paid generation task upstream while also incorrectly deducting user credits twice. A robust approach requires assigning a persistent local request ID before calling the provider and linking it to the provider's task ID only once a response is confirmed. A dedicated state model — distinguishing uncertain outcomes from confirmed failures — prevents premature credit releases or redundant submissions. Application credits should be reserved before any submission attempt and held until the outcome is definitively established, not released simply because a response was lost.