Seedream 5.0 Lite, measured: 1K returns a 400, and where a flat price helps
Seedream 5.0 Lite's distinguishing feature is a per-image price you can predict: raising resolution does not change it. It also failed to hold a referenced person. The use cases divide cleanly.
Measured
| Item | Measured value |
|---|---|
| Latency | ~41s |
| Cost per image | $0.035 (flat) |
| Output | JPEG |
| Measured on | 2026-08-28 (via OpenRouter) |
It was the slowest of the three. The flat unit price does make monthly budgeting straightforward, which is a real advantage at volume.
The first trap: resolution=1K fails
Nano Banana 2 accepts 512 upward, so a shared UI listing one set of resolution options will break on Seedream alone. Accepted values have to be held per model.
1K is small, so surely it goes through?
The opposite. The floor is 3,686,400 pixels, so it is rejected for being too small.
Accepted parameters
| Parameter | Accepted values |
|---|---|
| Aspect ratio | auto plus 18 options |
| Resolution | 2K / 4K (1K returns 400) |
| Images per run | 1–4 |
| Reference images | up to 14 |
Eighteen aspect ratios is the widest of the three models, which helps when you need an unusual frame.
Don't use it to name a person. For anything else, a predictable price is worth a lot.
It did not hold identity
Given a reference, it returned a different person. It does reproduce style and mood, so a reference works as a tonal instruction — but not as "this specific person".
This is a property of the model, not something prompting fixes. Where identity matters, choose a different model up front.
Good for
- Subjects without a person — backgrounds, objects, landscapes, abstracts
- Budgeting against a fixed per-image cost
- Unusual aspect ratios
Frequently asked
- Can Seedream 5.0 Lite output 1K images?
- Sending resolution=1K returns HTTP 400. A 3,686,400-pixel minimum applies, making 2K the floor.
- What does it cost?
- A flat $0.035 per image measured. Raising resolution does not change the unit price.
- Can it keep the same person?
- Not in our tests — a reference still produced a different person. Use GPT Image 2 when identity matters.
