Do reference images raise the price of a Sume image edit?

On most Sume image rows, adding reference images does not change the price: you pay the flat billed amount per output image. The exceptions are GPT Image 2, which adds $0.013 per input image, and GPT Image 2.5 and its variants, which are token metered, so an edit adds an estimate for its input tokens. Check the endpoint price before a batch of edits.

This is from the pricing code on origin/main and the Image API docs, read on 2026-10-10. The wallet charge is the endpoint cost_usd times n, and the docs say the endpoint lines are already billed amounts.

Three pricing shapes

A flat row has one price per output image, with no change for references. GPT Image 2 adds a fixed amount per input image. GPT Image 2.5 meters tokens: quality and size set the output cost, and the estimate for an edit includes the input images.

Vendors differ too. xAI's image generation guide, read the same day, says edits bill input and output images. Do not assume the vendor's billing rule carries over to Sume's row: Sume's catalog price for the Grok row is a flat $0.025 per output image.

Worked example: a product edit with three references

A product edit with three references on GPT Image 2 costs the base $0.26375 plus 3 x $0.013 = $0.039, for $0.30275. The same edit on FLUX.2 Pro is $0.0375, with no change for references. On GPT Image 2.5 the cost is the output price plus the input token estimate, so ask the endpoints route for the row and size you plan to use.

For a hundred such edits the three figures come to $30.28 on GPT Image 2, $3.75 on FLUX.2 Pro and an amount you read from the 2.5 pricing lines. The spread is a reason to estimate before you commit a team to one row.

How to estimate safely

Read the pricing lines from the endpoints route for the row, add the input-image term if the row has one, multiply by n, and compare to your wallet. If a number looks too low for a row with token metering, assume the estimate is conservative and watch the first few real charges.

Keep references to what the edit needs. Ten references on a token-metered row costs more than two, and on a flat row costs the same, so fewer is a free saving on one and a real saving on the other.

• Flat rows: references are free.
• GPT Image 2: $0.013 per input image.
• GPT Image 2.5: tokens, so estimate.
• Text-only rows: remove the references or switch rows.

A note on `n`

Because the charge is the endpoint price times n, four variants of an edit cost four times the single price on a flat row. On GPT Image 2, each of the four also adds its input-image term, so a three-reference, four-variant call costs four times the per-image total. Keep that in mind when you widen a test: the number of variants multiplies everything, and the number of references only multiplies on the metered rows.

If a call times out and becomes a job, the same price applies to the job. The docs describe polling the job rather than resending, which avoids paying twice for the same render.

The takeaway is simple. Before an edit-heavy project, price one representative edit on each candidate row, with the real number of references and the real size. Ten minutes of arithmetic tells you whether the project costs $40 or $400, and it is much easier to change rows before the batch than after.