Seedream 5.5: ByteDance's Editing Model Beats 5.0 Pro at Less Than Half the Price per Image
ByteDance released Seedream 5.5 on November 9, 2026: $0.020 to $0.035 per image, native transparent output, 99-layer decomposition, and stable multi-turn edits. Self-reported evals.

Key Numbers
Seedream 5.5 · Nov 9, 2026
Output at 2K
vs Seedream 5.0 Pro
ByteDance Seed released Seedream 5.5 on November 9, 2026: a production image model built for editing workflows rather than single shots, at less than half the per-image output price of Seedream 5.0 Pro. On BytePlus ModelArk it ships as Dola Seedream 5.5 (dola-seedream-5-5-261028).
The release is about workflow primitives: pixel-level preservation of unedited regions across long edit chains, native transparent (RGBA) output for every generation task type, and layer decomposition into up to 99 editable layers with text OCR. On ByteDance's self-reported evals it beats 5.0 Pro on every published comparison and leads GPT Image 2.5 Sunburst on all four GSB rows, with the lowest text-corruption rates in its chart (0.4% severe, 1.7% minor). Against Nano Banana 2.1 the picture is closer: small text-structure leads, a clear text-to-image realism lead, and one loss on image-to-image realism. Read the Seedream 5.5 launch blog and the BytePlus ModelArk docs.
What's New
ByteDance frames three headline upgrades over Seedream 5.0 Pro, which it calls the previous version.
Multi-turn editing
Seedream 5.5 is trained to preserve unedited regions at the pixel level, eliminate color shifts in those areas, and reduce detail loss across successive edits. ByteDance says this supports continued work on the same image across 100 rounds of editing. In its demo, ByteDance performed 100 successive edits on an image of the same person, using each round's output as the next input. That is a demonstration, not a measured metric.
Native RGBA
The model can directly generate images with an alpha channel, including full and partial transparency, with no separate background removal or edge cleanup. Unlike Seedream 5.0 Pro, which supported RGBA only for single-reference image-to-image with one PNG alpha input, Seedream 5.5 offers native transparent output on text-to-image, single-reference image-to-image, and multi-reference image-to-image.
Realism and text
ByteDance claims more natural lighting, materials, and skin, with less gloss and overexposure, plus more accurate dense small text and complex characters in English and Chinese.
API-level features
- Layer decomposition: 1 base image plus up to 99 layers (Seedream 5.0 Pro: 16). Configurable max output count from 2 to 100 including the base. Rich-text OCR returns text content, position, and approximate style for regular text layers (not artistic or handwritten text).
- References: up to 10 reference images for image-to-image. Inputs need no alpha channel.
- Background:auto (default), transparent, or opaque. For the most consistent multi-turn editing, ByteDance's docs tip is to use an RGB input and set background to opaque.
Editing workflow
100-round demo · 99 layers · RGBA
Round 1
edited · preserved at pixel level
Round 2
edited · preserved at pixel level
Round 3
edited · preserved at pixel level
...
continued
Round 100
edited · preserved at pixel level
Layer decomposition
1 base + up to 99 layers
text layer: OCR
Native RGBA
native RGBA
Transparent output for text-to-image and image-to-image
ByteDance also builds video models in the same Seed line. See our Seedance 2.5 launch post for the video counterpart.
Evaluations
Every evaluation number below is self-reported by ByteDance Seedin its November 9, 2026 launch blog, run on ByteDance's MagicArena platform with ByteDance's evaluators. These are not independent benchmarks. ByteDance does not publish the GSB formula or sample sizes.
Setup: six production use cases (e-commerce, marketing, film and television, gaming, office productivity, fashion design), rated by experienced image-generation evaluators. ByteDance says Seedream 5.5 improved both quality scores and usability rates across all six use cases, and that usability in film and television and in office productivity was more than twice the previous version. The usability and MOS radar chart has no value labels, so those remain qualitative claims only.
GSB preference for Seedream 5.5
0 = parity; positive favors Seedream 5.5
GPT Image 2.5 Sunburst
Nano Banana 2.1
Seedream 5.0 Pro
Text corruption rate (lower is better)
GSB preference (self-reported by ByteDance)
Positive GSB favors Seedream 5.5. GPT Image 2.5 Sunburst is OpenAI's previous generation; ByteDance did not test GPT Image 2.6 or Nano Banana Pro.
| Comparator | Text structure, I2I | Text structure, T2I | Realism, I2I | Realism, T2I |
|---|---|---|---|---|
| GPT Image 2.5 Sunburst | +20.72% | +16.01% | +36.05% | +37.29% |
| Nano Banana 2.1 | +3.82% | +5.23% | -6.18% | +21.79% |
| Seedream 5.0 Pro | +18.54% | +5.00% | +55.62% | +39.44% |
Text corruption rate (self-reported by ByteDance)
Lower is better. Chinese and English, including dense small text. By our arithmetic, Seedream 5.5's severe rate is one seventh of Seedream 5.0 Pro's (0.4% vs 2.8%).
| Model | Severe | Minor |
|---|---|---|
| Seedream 5.5 | 0.4% | 1.7% |
| Seedream 5.0 Pro | 2.8% | 2.6% |
| Nano Banana 2.1 | 1.7% | 3.0% |
| GPT Image 2.5 Sunburst | 0.7% | 7.4% |
Honest reading: decisive over Seedream 5.0 Pro and GPT Image 2.5 Sunburst, close to Nano Banana 2.1 on text structure, behind it on image-to-image realism, and no head-to-head with GPT Image 2.6 or Nano Banana Pro.
Pricing
Official BytePlus ModelArk list prices, fetched November 16, 2026. Seedream 5.5 bills every input image at $0.008. Output is $0.020 per image from 921,600 to about 2.60 million px (covers 1K and 1.5K) and $0.035 above that up to about 4.62 million px (2K). Default size is 2K, so a default call bills $0.035. Images blocked by moderation are not billed. In layer decomposition each layer is billed as an output image at its own tier. No Batch or discount tier is listed.
| Model | Input image | Output (1K to 1.5K) | Output (2K) |
|---|---|---|---|
| Seedream 5.5 | $0.008 | $0.020 | $0.035 |
| Seedream 5.0 Pro | First free, then $0.003 | $0.045 | $0.09 |
| Seedream 5.0 Flash | Free | $0.018 | $0.018 |
By our arithmetic on list prices: output vs 5.0 Pro is 55.6% lower at 1K to 1.5K and 61.1% lower at 2K. A one-reference edit at 2K costs $0.043 vs $0.09 on 5.0 Pro (52.2% lower); at 1K to 1.5K it is $0.028 vs $0.045 (37.8% lower). A 100-round edit chain at 2K with one input per round is $4.30 on 5.5 vs $9.00 on 5.0 Pro.
Output price per image
list prices · Nov 16, 2026
About 1K (1024x1024)
Nano Banana 2.1 Batch
$0.0168
Seedream 5.0 Flash
$0.018
Seedream 5.5
$0.02
Nano Banana 2.1
$0.0336
Seedream 5.0 Pro
$0.045
Nano Banana Pro
$0.134
2K (2048x2048)
Seedream 5.0 Flash
$0.018
Nano Banana 2.1 Batch
$0.0252
Seedream 5.5
$0.035
Nano Banana 2.1
$0.0504
Seedream 5.0 Pro
$0.09
Nano Banana Pro
$0.134
Versus Nano Banana 2.1 Standard, Seedream 5.5 output is 40.5% lower at 1K ($0.020 vs $0.0336) and 30.6% lower at 2K ($0.035 vs $0.0504). Nano Banana 2.1 Batch at 2K ($0.0252) is cheaper than Seedream 5.5 at 2K. Nano Banana Pro at 2K ($0.134) is 3.8 times Seedream 5.5. Caveat: Seedream bills by pixel tier and Google by tokens per resolution; this compares output prices for a 1024x1024 or 2048x2048 image only. Input costs are structured differently. Sources: BytePlus ModelArk pricing and Gemini API pricing.
OpenAI prices GPT Image 2.5 Sunburst and GPT Image 2.6 by token ($30 per 1M image output tokens, $8 per 1M image input tokens; Batch $15 and $4) and publishes no per-image figure, so those models are not on the rails. See OpenAI pricing.
Specs
| Detail | Value |
|---|---|
| Model id | dola-seedream-5-5-261028 |
| Inputs | Text and images (up to 10 references) |
| Output | Image (png default, or jpeg) |
| Tasks | Text-to-image, single-ref and multi-ref image-to-image, interactive editing, layer decomposition |
| Resolutions | 1K, 1.5K, 2K (default 2K). No 4K. |
| Max pixels | About 4.62 million (custom size from 1280x720) |
| Aspect range | 1:16 to 16:1 |
| Background | auto (default), transparent, opaque |
| Layers | 1 base + up to 99 layers; OCR on regular text layers |
| Watermark default | On ("AI-generated" mark, lower right) |
| Rate limit | 500 images per minute |
| Prompt languages | English, Chinese, plus Russian, Arabic, Filipino, Thai, Turkish, Korean, Malay, Spanish, Portuguese, Indonesian, French, German, Vietnamese, Japanese |
| Availability | BytePlus ModelArk; rolling out on Jimeng, Doubao, Volcano Engine Ark. Proprietary, API only. |
| Not published | Parameters, architecture, training data, knowledge cutoff, latency |
More detail: BytePlus image generation API and the Seedream 5.5 project page.
Who Should Use It
- Pick Seedream 5.5for iterative edit pipelines, transparent assets (stickers, icons, product cutouts), layered design files with editable text, and dense small text in Chinese or English. It replaces Seedream 5.0 Pro for most work: better on ByteDance's evals and cheaper.
- Pick Seedream 5.0 Flash for cheap single generations ($0.018 flat, free inputs) when you need none of the 5.5 features.
- Pick Nano Banana 2.1if you need 4K, a Batch discount, or the strongest image-to-image realism in ByteDance's own chart.
- There is no head-to-head evidence vs GPT Image 2.6 or Nano Banana Pro; test on your own prompts.
- Waiting for Seedream 5.5 Pro is reasonable if you want ByteDance's top tier; it has no date or price.
Banner: ByteDance Seed.
FAQ
- ByteDance released Seedream 5.5 on November 9, 2026. On BytePlus ModelArk it is listed as Dola Seedream 5.5 with API id
dola-seedream-5-5-261028. ByteDance's Chinese announcement says it is rolling out on Jimeng, Doubao, and Volcano Engine Ark. - Output is $0.020 per image from about 1K to 1.5K and $0.035 at 2K. Every input image costs $0.008. Seedream 5.0 Pro charges $0.045 and $0.09 per output image (first input free, then $0.003). By our arithmetic on list prices, Seedream 5.5 output is 55.6% lower at 1K to 1.5K and 61.1% lower at 2K. There is no Batch tier.
- Pixel-level preservation of unedited regions across long edit chains, native RGBA for text-to-image and image-to-image (5.0 Pro only supported RGBA for single-reference image-to-image with one PNG alpha input), layer decomposition into up to 99 layers instead of 16 with rich-text OCR, png as the default output format, and a lower per-image list price.
- ByteDance demonstrated 100 successive edits on one image and says the model supports continued work across 100 rounds. That is a demo, not an independent measured metric.
- Yes. Seedream 5.5 can directly generate images with an alpha channel for text-to-image and image-to-image. Set the background parameter to
transparent. Inputs do not need an alpha channel. Seedream 5.5 Standard output is cheaper per image at 1K and 2K. On ByteDance's self-reported evals it shows small text-structure leads, a clear text-to-image realism lead, and a loss on image-to-image realism (-6.18% GSB). Nano Banana 2.1 offers 4K and a Batch discount; Seedream 5.5 tops out at 2K and has no Batch tier.
- Presets are 1K, 1.5K, and 2K (default 2K). Custom sizes go from 1280x720 (921,600 px) to about 4.62 million pixels. Aspect ratio ranges from 1:16 to 16:1. There is no 4K tier.
- On the BytePlus ModelArk API as
dola-seedream-5-5-261028. ByteDance says it is also rolling out in ByteDance apps in China (Jimeng, Doubao, Volcano Engine Ark). It is proprietary and API only. ByteDance says a Pro version of the Seedream 5.5 series will follow. No date or price has been published.
Continue Reading
- Qwen4-Max: Near-Frontier Agent Work at $1.50 and $4.50
- Hy4 flash: Tencent's $0.07 Model Undercuts Haiku 5.5 and GPT-6 Luna
- Gemini 4 Hydrogen: Argon-Class Agent Work at $1.50 and $7.50, Flat to 1M
- Muse Ember: Meta's Volume Model at the Same $0.10 and $0.50 as Haiku 5.5
- Claude Mythos 5.2: Same Model as Fable, Different Guest List
