Hy Image 3.5 Preview 20 references vs GPT Image 2.5 16: which edits?

For a multi-image edit today, ChatGPT Image 2.5 is the one you can call on Sume, with up to 16 references and an optional mask_url; Tencent Hy Image 3.5 Preview lists up to 20 references on its OpenRouter page but is not in Sume's catalog. If your brief needs more than 16 inputs, Hy is the only one of the two that lists the room, and you would call it through its own listing.

The difference of four references only matters for sheets such as a lookbook or a character board. For a normal product edit with one photo and one logo, both are far above what you need.

Side by side

The Sume column comes from the Image API docs. The Hy column comes from the OpenRouter listing read today. Cells that a source does not state are marked as not stated, rather than filled in.

What a reference actually costs on Sume

On ChatGPT Image 2.5, references are input image tokens at $8 per million before Sume pricing, per the Fal token rates the docs cite. Input token counts are estimates, so each added reference raises the estimate and the more you send the more the call costs. Send only the images that change the result.

• Put the image to edit first and style or identity references after it.
• Use mask_url when only one region should change.
• Keep reference URLs public HTTPS; Sume rejects private, localhost and non-HTTPS URLs before submission.

If you have 17 to 20 inputs

More references only help if the model uses them. On Sume the ceiling is 16 per call, so a 17-to-20 image brief has to be reduced or split into more than one call. Price both options before you choose, since each call is a separate generation.

• Rank references by how much each changes the result.
• Drop near-duplicates before counting.
• Compare the result with a single call on one real brief.

How to choose

If the edit is a swap inside one frame, choose the model with a mask and a price you can read in the catalog. If it is a collage of many inputs where a 17th reference matters, test Hy directly and measure its token use, because the listing gives no per-image figure.

For a ChatGPT Image 2.5 reference cost grid with the input token math, read the 16 references versus one reference post.

The reference count is a ceiling, not a target. Run the same brief on both models with the same few references, compare the results, and add the cost of failed attempts before you decide.