Seedream 4 or Seedream 4.5 on Sume: 3.25 vs 5 cents per image
Seedream 4 bills $0.0325 and lists aspect ratio auto. Seedream 4.5 bills $0.05 and does not. Same n, references and formats. What the extra 1.75 cents buys.
On Sume, Seedream 4 bills $0.0325 per image and Seedream 4.5 bills $0.05, a 54 percent premium for the newer row. The catalog descriptors are otherwise almost the same: both take 1 to 4 images per call, up to 10 references, and png, jpeg or webp output. The one parameter difference in the contract is that Seedream 4 lists aspect_ratio: "auto" and Seedream 4.5 does not.
The catalog does not say what the extra 1.75 cents buys in picture quality, and this post does not claim it. It is a spec comparison taken from the Sume catalog on origin/main, read 2026-10-08. Test both on your own prompts before you decide.
What the catalog lists
Model ids are bytedance-seed/seedream-4 and bytedance-seed/seedream-4.5. Neither row lists resolution or quality, so a request that sets either returns 400 unsupported_parameter. Both are served through the single sume endpoint, so there is no provider choice to make.
| Field | Seedream 4 | Seedream 4.5 |
|---|---|---|
| Billed per image | $0.0325 | $0.05 |
| 100 images | $3.25 | $5.00 |
| Images per call (n) | 1 to 4 | 1 to 4 |
| Reference images | Up to 10 | Up to 10 |
| Aspect ratios | 1:1, 16:9, 9:16, 4:3, 3:4, 4:5, 5:4, 3:2, 2:3, plus auto | The same nine, no auto |
| resolution / quality | Not listed | Not listed |
| Output formats | png, jpeg, webp | png, jpeg, webp |
Why auto matters for edits
Docs say that on edit calls, aspect_ratio: "auto" matches the reference, and that omitting the field is not the same as auto. On Seedream 4 you can ask for the reference shape explicitly. On Seedream 4.5, you must name a ratio from the nine, so a reference with an unusual shape has to be matched to the nearest listed one, or you accept what the row returns when the field is left out.
If your edits start from photos in many shapes, this is the practical reason to keep Seedream 4 in a pipeline. If your sources are all 1:1, 4:5 or 16:9, the difference does not show.
A split that costs little
Send the exploratory batch to Seedream 4, where 100 candidates cost $3.25, and rerun the ones you keep on Seedream 4.5 only if side-by-side tests show it wins on your content. Ten keepers rerun on the newer row cost $0.50, so the whole loop is $3.75 for 100 candidates and ten finals.
Both rows reject seed, so a rerun is a new sample, not a repeat of the first image. Read usage.cost from each response and keep the total. The charge is the endpoint's cost_usd times n, and failed generations are not billed.
Keep the model id in configuration, not in code, so you can switch rows without a deploy. Log the id, the aspect ratio and usage.cost with each result. After a week you will have a small data set of your own showing which row produced the files you kept, and that is better evidence than any catalog line.
Sources
Related posts
More in Comparisons
- Speech-to-text per hour: Sume $0.60 batch vs MAI streaming $0.54
Sume STT is batch at $0.01 per audio minute, $0.60 an hour. MAI-Transcribe-2-Streaming lists $0.54 an hour through year end. Different products; the math.
- Sume Format vs pasting the same prompt each time: what changes
A Format stores the house style once, then each API run sends only inputs. How the instruction is composed, what a run returns, and the spend cap.
- Swapping a face in a 12-second clip: manual steps vs one request
What changing a face in a short clip involves when you do it yourself, and what Sume's beta face-swap request replaces. Fields, limits and the cost ceiling.
- Synthesia's 60 free Interactive Avatar minutes: what they cover
Every Synthesia plan now includes 60 one-time Interactive Avatar minutes at signup. What that buys, what it does not, and the rendered-clip route on Sume.
Written by Sume