Hy Image 3.5 at $0.024: which Sume image rows bill less?

Tencent Hy's launch post (read 2026-10-08) prices Hy Image 3.5 Preview at $0.024 per image on the Tencent Cloud API. On Sume, two rows bill less than that: Higgsfield Soul at $0.005 (720p) and ChatGPT Image 2.5 at low quality and a small size, whose fal floor of $0.00402 becomes about $0.005 at Sume's multiplier. Three rows sit a tenth of a cent above it at $0.025: Grok Imagine, Qwen Image and Imagen 4 Fast.

Sume does not list Hy Image, and the OpenRouter listing for it prices tokens, not images, so the two figures are not the same kind of price. The comparison is a ceiling to beat, not like for like.

The comparison

Sume bills each image at its provider list price times 1.25, rounded once in micro-dollars; the endpoint pricing line on the catalog carries the result. The figures below are from the repo's price tables on origin/main.

What the cheaper rows give up

Soul is text-to-image only, takes 1 or 4 images, and lists seven aspect ratios. ChatGPT Image 2.5 at low quality is token-billed, so a longer prompt or reference images add input tokens, and the floor applies only to the smallest size at low quality. Hy Image lists up to 20 references on OpenRouter; Sume's reference caps are 16 on ChatGPT Image 2.5, 10 on most rows and 5 on Ideogram 4.5.

A price per call also hides retries. Sume bills a completed generation in full and does not bill a failed one, so the cost of a good image is the sum of your usage costs divided by the images you keep.

How to check

Do not copy these numbers into a budget. Read GET /v1/images/models/{model_id}/endpoints for the row you plan to use and multiply cost_usd by n.

Price is a weak filter for a draft batch, because the cost of a bad draft is the time spent reading it. Pick the cheap row whose request shape matches the shot you need: a reference photo rules out Imagen 4 Fast, a 21:9 frame rules out Grok Image, and a text prompt with a standard ratio leaves all three open.

Run the same ten prompts on each candidate and total usage.cost from the responses. At these prices the whole comparison is well under a dollar, which is cheaper than arguing about it.