A misspelled Ideogram 4.5 edit still bills: budget for retries

If an Ideogram 4.5 edit on Sume returns an image with a misspelled word, you pay for it. Sume's Image API docs say billing is all-or-nothing: a generation that completes is billed in full, and a generation that fails or is cancelled is not billed. A wrong word in a good-looking image is still a completed generation, because the API cannot read the text it produced.

This matters more for text edits than for plain generation, since the whole point of the call is a few exact characters. Ideogram's launch post (read 2026-10-05) says 4.5 removes artifact buildup and makes multi-turn editing possible. That helps when you iterate. It does not make any single pass free.

What is and is not refunded

A worked example with symbols, not prices

Say you edit 40 product labels and your pilot shows that 1 in 5 first attempts has a wrong letter. With a single re-run allowed per label, the expected count of billed images is the 40 first attempts plus about 8 re-runs, so about 48. If the re-runs also fail at 1 in 5, add roughly 1.6 more. The point is the shape of the sum, not these figures: they are an example, not a measured rate for any model.

Multiply the expected image count by the per-image line from the catalog. The pricing array on the endpoint record lists billable lines such as output_image with a cost_usd; the Image API docs show that shape and say you pay cost_usd x n.

Put a gate between the draft and the final

Because wrong text is billed, the cheapest control is a human or scripted check at the low tier before you pay for the high tier. A person reading 40 labels takes a few minutes. A mistake found after a high run costs a high-tier image. Gate every batch the same way: draft, read, then promote only the good prompts.

A retry budget

Pick the quality tier first. The docs list Fal list prices of $0.03, $0.06 and $0.22 per image for low, medium and high, with Sume's lines including its margin. Then decide how many attempts per image you will allow, and multiply:

Budget = images x (1 + expected re-runs) x the per-image line from GET /v1/images/models/ideogram/ideogram-v4.5/endpoints.

Track your own re-run rate on a pilot of 20 images before you scale. Do not assume a number from a vendor page; no vendor figure for text-edit accuracy was read for this post.

Cheaper ways to retry

Run drafts at quality: low and promote only the prompts that read correctly to a higher tier. Put the exact string in quotes, and keep the new text close to the old text in length. When you retry a call that failed, send the same Idempotency-Key so a retry cannot create a second job. When you retry a call that completed with wrong text, send a new key, because it is a different request: the errors page says to use the same key only for the same operation and payload.