Seedream 5.5 is ByteDance's image model, released November 9, 2026. Its main change is editing: in our test, after ten edits in a row the untouched parts of a picture had drifted less than they did on Seedream 5.0 Flash after one. It also makes transparent pictures from a prompt, and costs 4 credits a picture in Tonpit, against 9 for Seedream 5.0 Pro.
Tested November 12, 2026: 82 pictures, the same prompts on Seedream 5.5, Seedream 5.0 Pro and Seedream 5.0 Flash from BytePlus, and GPT Image 2.6 Flare, at the size Tonpit sells each model. They cost $2.15 at the suppliers' rates. Prices below are in credits as Tonpit charges them; a credit is one cent. Tonpit lists Seedream 5.0 Pro as Seedream 5.
What is Seedream 5.5?
- A model for pictures and edits. It makes a picture from text, or edits one from up to 10 input pictures. One picture per request, no seed.
- Three sizes: 1K, 1.5K and 2K. There is no 4K. Every size we ordered came back at exactly that size.
- A transparent background from a prompt alone or from ordinary pictures. Seedream 5.0 needed an input that was already transparent.
- It rewrites your prompt before it draws, and that cannot be switched off.
- It can split a picture into layers. Tonpit does not offer that yet; we tried it once, below.
How does Seedream 5.5 compare to Seedream 5.0 and GPT Image 2.6?
On a single edit it moved the rest of the picture least of the four models, and its pictures read as plain photographs where Flash goes for drama and GPT Image 2.6 looks polished. On text it was level with them. Each row is one prompt run on the models named.
| Test | Seedream 5.5 | Seedream 5.0 Flash | GPT Image 2.6 |
|---|---|---|---|
| Fisherman portrait | Muted, like a plain photograph | Dramatic, an odd angle | Polished, posed |
| Bottle with a 3-line label | Spelled right | Spelled right | Spelled right |
| Chalkboard menu, 8 lines | Spelled right | Spelled right | Spelled right, at High |
| Jar back label, small print | Spelled right | Spelled right | Spelled right, at High |
| One edit: drift of the rest | 0.48 | 3.75 | 4.24, at Low |
| Six edits in a row: drift | 2.65 | 15.14 | Not run |
| Transparent from a prompt | Real transparency | Not offered in Tonpit | Real transparency, finer hair |
| Two people from references | Both kept | Both kept | Not run |
How well does Seedream 5.5 hold a picture across many edits?
After ten edits in a row, the parts of the picture nobody asked to change had drifted less than they did on Seedream 5.0 Flash after one. Flash is the only other model we ran the whole chain on. We edited one room six times, each edit on the last one's result: tulips for the lamp, an apple, a cat, a new print, an open laptop, a pair of shoes. Then four more on 5.5.
Before any edit ran we marked what no edit touches: the cork board, the dresser, the bookshelf and two window panes. After each round we measured how far those parts had moved from the original, as colour difference (ΔE). Under 1 the eye sees nothing, around 2 to 3 a trained eye notices, and past 10 anyone does.
| After | Seedream 5.5 | Seedream 5.0 Flash |
|---|---|---|
| 1 edit | 0.49 | 3.91 |
| 3 edits | 0.94 | 8.25 |
| 6 edits | 2.65 | 15.14 |
| 10 edits | 3.34 | Not run |
All six edits landed on both models, so 5.5 does not hold the picture by ignoring the request. On a single edit, Seedream 5.0 Pro drifted 2.94 and GPT Image 2.6 at Low 4.24, against 0.48. Three cautions: this is one chain per model, the drift on 5.5 jumped at round six for a reason we did not find, and ByteDance's claim is 100 rounds, where we ran ten.
Can Seedream 5.5 make a transparent background?
Yes, from a prompt or from pictures you give it: seven of seven transparent orders came back with real transparency, soft edges included. That covers a sticker, a glass bottle, hair, smoke, a person lifted out of a photograph, and four subjects taken from four separate pictures and set side by side.
GPT Image 2.6 makes real transparency too, at the same 4 credits on a wide frame, and its loose hair was finer. Seedream 5.5 drew a soft shadow we asked for as a pale haze, which shows as a glow on a dark ground. Neither model made glass see-through.
What does Seedream 5.5 cost, and how fast is it?
4 credits a picture at 1K and 1.5K and 7 at 2K, and every picture you give it adds to the bill: an edit of one picture is 6. That is less than half of Seedream 5.0 Pro. Flash stays cheaper once you feed it pictures or order 2K, because its price is flat.
| Model | A wide picture | The same at 2K | An edit of one picture |
|---|---|---|---|
| Seedream 5.5 | 4 | 7 | 6 |
| Seedream 5.0 Flash | 4 | 4 | 4 |
| Seedream 5.0 Pro | 9 | 18 | 9 |
| GPT Image 2.6 Flare, Low | 4 | No 2K; 10 at 4K | 7 |
Faster than Seedream 5.0 Pro, not faster than Flash. A picture took 10 to 22 seconds on Seedream 5.5, 13 to 18 on Flash, and 36 to 46 on Seedream 5.0 Pro, whose one edit took 101. GPT Image 2.6 at Low was the quickest, at 6 to 9.
Can it keep the same person across pictures?
Yes. We gave it one small portrait of Rúna, one of Tonpit's ready-made fictional characters, and asked for three new scenes. Her face, red braids and patterned sweater held in all three. Each of these cost 6 credits, because the portrait is an input.
What is layer decomposition in Seedream 5.5?
It splits one flat picture into a base and separate transparent layers, and reads out the text. We gave it a poster with a title, two lines of small text and three objects. In 30 seconds it returned the cream base and seven layers, each with its position, and for the text layers the words, a nearest font and a colour. It also lifted the lettering on the food cart as a layer nobody asked for.
Every layer bills as a picture, so that one request cost what eight pictures cost. Tonpit does not offer layers yet. For a video studio it is the part of this model with the most to give: a flat picture becomes parts that can move.
Which Seedream model should you use?
Seedream 5.5 for a picture you will edit more than once, and for a transparent one. Flash when you give the model several pictures or want 2K at the lowest price. Seedream 5.0 Pro has one job left in Tonpit: a face that Seedance has to accept, since BytePlus's list of sources Seedance trusts for faces names the 5.0 models and not 5.5.
What to know before you rely on it. The prompt rewrite cannot be turned off, and some pictures came back with things we did not ask for: a brand name on a perfume bottle, Chinese signs in a market. A transparent order clears the backdrop, not the scene: a paper crane we asked for alone came back on a table among other cranes, until we named what must not be there. In Tonpit, Seedream models from BytePlus are not available in the United States, where GPT Image 2.6 covers the same ground. And ByteDance has said a Pro version of 5.5 will follow.