ByteDance Seed's image model, released November 9, 2026 — and this site's second image-model topic: 100 rounds of editing on the same image with pixel-level preservation of untouched areas, native transparent PNG output with an alpha channel, and a vendor-reported severe text-breakage rate of just 0.4%. But the 100-round demo covers a single portrait, the longest independent test is 18 rounds, and glass and water come out opaque in hands-on tests. No official API price at launch; Volcano Ark published its price list on Nov 12 (a 2K edit costs about ¥0.28), while BytePlus still has no confirmed per-image price as of Nov 14. Independent scores arrived Nov 11: AA text-to-image 1135, editing 1159 — both below GPT Image 2.6 Sunburst. This page collects official specs, pricing, API status and first readings, and marks what is still unverified.
The confirmed picture first: Seedream 5.5's 100-round editing and transparent PNG are real capabilities, but the former was demoed to 100 rounds on one portrait only, and the latter fails on glass and water in hands-on tests; pricing is public in China and still pending internationally; independent scores are out — good, but not first.
Doubao-Seedream-5.5 on Volcano ArkSeedream went from 4.5 to 5.5 in under a year; within four days of the 5.5 launch, independent scores, third-party APIs and the Volcano Ark price list all landed — but BytePlus's official international API still has no price.
Seedream 4.5 released.
Seedream 5.0 Lite released (83 days after 4.5).
Seedream 5.0 Pro released (134 days after 5.0 Lite), with the layer-decomposition endpoint (up to 16 layers) the same day.
GPT Image 2.6 released and added to Arena — five days before 5.5; ByteDance's comparison copy names the older GPT Image 2.5 Sunburst (see section 09).
Seed official blog + X announce Seedream 5.5; Dreamina, Doubao, Doubao Work and the Volcano Ark experience center go live the same day (unlike 5.0 Pro, which came to Ark first and apps later, all four opened at once).
WaveSpeed lists three 5.5 endpoints the same day (text-to-image, editing, layer decomposition); Quantslant runs a 35-image test on WaveSpeed that night.
Doubao announces its image-design upgrade with 5.5; BytePlus launches its product page and image packs; ComfyUI v0.45.0 ships with a Seedream 5.5 Partner Node.
ComfyUI blog announces Partner Nodes; Pika Image Studio adds 5.5; Artificial Analysis image boards list 5.5.
OpenRouter lists it (page says "released November 12" — the listing date); Seedance.io runs a same-prompt three-model test (38 requests); the Volcano Ark price list is publicly readable.
Ofox lists it (model ID volcengine/doubao-seedream-5.5); Rapidata dataset syncs, 5.5 overall Elo 1103 (#10).
This topic page goes live; BytePlus still has no confirmed official per-image price or ModelArk model ID.
ByteDance published no spec sheet; the figures below come from the official blog, third-party API docs and hands-on tests (checked Nov 10–14, 2026). Note: 2K maximum, no 4K; architecture and parameters are unpublished.
Doubao-Seedream-5.5; BytePlus ModelArk model ID unconfirmed as of Nov 14 (nowrap reports Dola-Seedream-5.5 listed, no price)Fastest in China: Dreamina, Doubao and the Volcano Ark experience center went live on launch day; internationally BytePlus, Dreamina and CapCut are rolling out gradually, account by account; third-party APIs already cover everything.
Dreamina web (Create → Generate/Canvas → Seedream 5.5), Doubao app/web (image generation → Seedream 5.5), Doubao Work (skills entry). Dreamina ran a launch promo: unlimited 1.5K for 7 days (non-members) / 30 days (members) — single source, a limited-time offer that may have ended; check the current Dreamina page. Some users reported on day one that it was flagged as a members-only model.
exp.volcengine.com: visual models → image generation → Doubao-Seedream-5.5. The API was promised "this week" at launch; the price list went public Nov 12 (see section 05). This is the only channel with a confirmed official per-image price and model ID.
Dreamina, CapCut and BytePlus are rolling out gradually, account by account (the Dreamina model page is live but access depends on the account). BytePlus sells image packs: Starter 700 images for $6.99 (first purchase, 30 days, ~$0.01/image). No ModelArk model ID or per-image price confirmed as of Nov 14.
WaveSpeed (three endpoints on launch night), OpenRouter (bytedance-seed/seedream-5-5), Atlas Cloud, Kie, Siray, SeedRouter, Segmind, apimodels, Ofox (volcengine/doubao-seedream-5.5). All reseller prices, not ByteDance pricing, and they change. Note: TeamoRouter still hadn't added it as of Nov 13, and OrcaRouter says it offers no ByteDance image models.
Higgsfield web: pick Transparent in the Background menu, 1–4 images per run, 2K max. Pika Image Studio added 5.5 on Nov 11 (100+ layer decomposition for fine editing). SeedVideo AI listed it Nov 12; new accounts get free credits usable on 5.5 (a reseller perk, not an official free tier).
v0.45.0 (Nov 10) includes a Seedream 5.5 Partner Node, announced Nov 11: cloud billing, no local weights (5.5 has no public weights to download). Estimated from the node source formula: 1K/1.5K ~$0.0286, 2K ~$0.05, ~+$0.011 per reference image.
China's official price list went public Nov 12: two output tiers by pixel count plus a per-input-image fee, a 2K edit for about ¥0.28; BytePlus internationally sells only image packs, no per-image price; third-party resellers range from $0.017 to $0.05.
| Item | Seedream 5.5 | Seedream 5.0 Pro | Seedream 5.0 Lite |
|---|---|---|---|
| Output ≤2.61MP (1K/1.5K tier) | ¥0.13 / image | ¥0.30 / image | ¥0.22 / image |
| Output >2.61MP (2K tier) | ¥0.23 / image | ¥0.60 / image | ¥0.22 / image |
| Input image | ¥0.05 each | First free, then ¥0.02 | Free |
| Text tokens | Free | Free | Free |
| 2K edit (1 input image) | ~¥0.28 | ~¥0.60 | ~¥0.22 |
5.5's 2K output is about 62% cheaper than 5.0 Pro (the vendor says "60% cheaper"); but 5.5 charges for every input image, so multi-reference edits narrow the gap. ¥0.28 is the figure quoted across launch coverage (¥0.23 output + ¥0.05 input).
| Platform | 1K / 1.5K | 2K | Per input image |
|---|---|---|---|
| WaveSpeed | $0.02 | $0.035 | +$0.008 |
| OpenRouter | $0.02 | $0.035 | +$0.008 |
| SeedRouter | $0.017 | $0.02975 | +$0.0068 (after first) |
| Kie | $0.018 | $0.0315 | — |
| Siray | $0.02 | $0.028 | — |
| ComfyUI Partner Node | ~$0.0286 | ~$0.05 | ~+$0.011 |
| apimodels | $0.03 | $0.05 | +$0.01 |
WaveSpeed's 2K price was $0.027 on Nov 9 and rose to $0.035 on Nov 10; layer decomposition from $0.60. Artificial Analysis measured costs: text-to-image $19.50/1,000 images, editing $27.00/1,000.
As of Nov 14, BytePlus's price list and ModelArk API reference show no 5.5 per-image price or model ID (last third-party reads Nov 11–12; nowrap reported Nov 13 that ModelArk lists Dola-Seedream-5.5 as available, with no price). The product page sells image packs: Starter 700 images for $6.99 (first purchase, 30 days, ~$0.01/image), plus 1,800 images at 31% off and 3,500 at 29% off. For reference, 5.0 Pro on BytePlus: ≤2.61MP $0.045/image, >2.61MP $0.09/image.
Official pages: Seed official blog · BytePlus Seedream product page · Volcano Ark experience center
Five days of testing circles three questions: how stable is 100-round editing really, is the transparent PNG real, and does text actually hold up. The answers so far: editing is much steadier than the previous generation, cutouts are clean, but glass stays opaque; text looks good in small samples — and the samples are all small.
Nov 9 night, 35 images ($0.77): Polish text — 10 errors across 823 characters in 11 images (6 images error-free); against GPT-Image-2.6 Sunburst (high) on the same 5 prompts: 3 errors vs 1 in 389 characters, costing $0.114 vs $0.214. Second round Nov 10: transparency 10/10 genuine RGBA PNGs, but glass and water 0/4 see-through, smoke 2/2 genuinely semi-transparent, hair and mesh bags cut cleanly; 27/30 Polish short headlines error-free (Nano Banana 2.1: 30/30). Also: a dense newspaper front page had 4 errors and hallucinated a real radio station's logo.
Same-prompt three-model test: on a poster with small text, 5.5 spelled everything right with the requested casing, 5.0 Pro got it all right but all-caps, 5.0 Lite misspelled two words. Six-round edit drift (2K output): wall-color drift 3.5 for 5.5 vs 17.3 for 5.0 Pro (0–441 scale); untouched bottle-cap area changed 0.6–1.7 vs 3.6–3.7 per pixel per round (0–255 scale) — 5.5 is steadier, but the label paper still darkened slightly after 6 rounds. On speed: text-to-image 2K median 47s for 5.5 vs 43s for 5.0 Pro (5.5 is not faster); single-reference edit 2K median 61s vs 97s (about 38% faster, matching the vendor's "nearly 40%").
A five-person poster went through 18 rounds of editing from modern styling to a period mystery look, with faces and composition holding up well — the longest independent multi-round test we found; 18 rounds does not verify 100. Also: Mona Lisa outfit and profession changes, a bare room gradually furnished (the view and sunlight angle outside the window unchanged). Layer decomposition came out at about 100 layers, with structured recognition results for text layers. The first transparency test still returned a normal background — an explicit alpha-channel request was needed.
The GPT Image 2.5 Sunburst ByteDance names joined Arena on Sep 8; its successor GPT-Image-2.6 Sunburst joined Nov 4, five days before 5.5 launched; but the blog says "leading models such as GPT Image 2.5 Sunburst and Nano Banana 2.1" — it never says the comparison was limited to those two, so whether 2.6 was included can't be confirmed, and there's no basis to say ByteDance deliberately picked the older generation. Also noted: the 100-round demo covers one portrait, and the blog's wording stops at "among the best" — it never claims first place. Cited the Banana100 study: repeated editing accumulates artifacts, and none of 21 image-quality metrics reliably identifies degraded images.
The AA editing board doesn't test repeated multi-round editing — the 100-round claim and the leaderboard score are two orthogonal questions: a high board score doesn't prove 100 rounds, and the 100-round demo can't be derived from board scores. Judge them separately.
Six in-depth articles from this site (Chinese + English), one question each, rewritten from public sources with a reference link.
On launch day (Nov 9) neither AA board nor Design Arena listed 5.5; from Nov 11 AA lists it: text-to-image 1135 (#10), editing 1159 (#6). Below GPT Image 2.6 Sunburst, statistically tied with Nano Banana 2.1 on editing. Scores are Nov 11–12 snapshots and move with votes.
| Model | Elo | Rank | Note |
|---|---|---|---|
| GPT Image 2.6 Sunburst (max) | 1261 | — | 126 points above 5.5 |
| Nano Banana 2.1 | 1176 | — | 41 points above 5.5 |
| MAI-Image-2.6 | 1170 | — | Microsoft image model |
| Seedream 5.5 | 1135 | #10 (of 109 models) | 5,128 votes; statistically tied with #9–#13 |
| Seedream 5.0 Pro | 1102 | — | Previous flagship, 33 points below 5.5 |
| Model | Elo | Rank | Note |
|---|---|---|---|
| GPT Image 2.6 Sunburst (max) | 1209–1210 | — | About 50 points above 5.5 |
| Seedream 5.5 | 1159 | #6 | 4,901 votes, 95% CI 1145–1173 |
| Nano Banana 2.1 | 1152 | — | Confidence intervals overlap — a statistical tie |
| Seedream 5.0 Pro | 1129 | — | 30 points below 5.5 |
7,679,896 human feedback entries, 57 text-to-image models, 1,514 prompts: Seedream 5.5 overall Elo 1103 (#10), with Preference 1200 (#6), Coherence 1146 (#4), Alignment 1068 (#16), Text coherence 1015 (#27). For reference: Seedream 5.0 Pro overall 965 (#30), 5.0 Lite 981 (#24), 4.5 1016 (#20).
The AA editing board doesn't test repeated multi-round editing — the 100-round claim and the board scores are orthogonal (OrcaRouter). No GenEval, DPG-Bench or other automated benchmark scores exist; the vendor itself publishes no evaluation table beyond MagicArena/GSB. Whether Design Arena / LMArena image boards have listed 5.5 after Nov 9 is unconfirmed in any source as of Nov 14.
On AA scores 5.5 is not first: text-to-image and editing both trail GPT Image 2.6 Sunburst, and editing ties Nano Banana 2.1. Its edge is elsewhere: multi-round editing stability, native transparent PNG, layer decomposition, and China-side pricing.
| Dimension | Seedream 5.5 | GPT Image 2.6 Sunburst | Seedream 5.0 Pro |
|---|---|---|---|
| AA text-to-image | 1135 (#10) | 1261 | 1102 |
| AA editing | 1159 (#6) | 1209–1210 | 1129 |
| Max resolution | 2K (no 4K) | 1K–4K (4K confirmed by third-party tests; guide doesn't list a 2.6-specific ceiling) | 2K |
| Multi-round editing | 100-round claim (single-portrait demo); longest independent test 18 rounds | Markup layers + saved reference images | Pixel-level editing, no round claim |
| Transparent background | Native RGBA, transparent PNG from text alone | Transparent backgrounds supported | Can only preserve existing transparency |
| Unique capability | Layer decomposition, ~100 layers | Native GIF (2–16 frames, default 9), SVG export, complexity control | — |
| Billing | Volcano Ark ¥0.13/¥0.23 + ¥0.05 per input image | Token-based (image output $30/1M tokens) | Volcano Ark ¥0.30/¥0.60 |
| Weights | Not public | Not public | Not public |
Nano Banana 2.1 (Google, joined Arena Oct 6): AA text-to-image 1176, editing 1152 — ahead of 5.5 on text-to-image, statistically tied on editing. ByteDance claims a lead over it on text structure (vendor evaluation, not independently confirmed). Quantslant's Polish short-headline test: Nano Banana 2.1 was 30/30 error-free, 5.5 was 27/30. Microsoft's MAI-Image-2.6 scores 1170 on AA text-to-image, above 5.5. Black Forest Labs' October release FLUX 3 Image also claims multi-step preservation of non-target areas, with no round count given.
MaverickFrame and BeatAPI reach the same conclusion: long feedback chains + transparent assets + dense Chinese text → 5.5; GIF/SVG/precise point-and-edit → 2.6. See our Seedream 5.5 vs GPT Image 2.6 article and our GPT Image 2.6 topic page.
Five days after launch, these are still unknown: some await ByteDance, some can only come from independent reproduction.
The confusion around 5.5 comes from four places: the evidence range for 100 rounds, the 4K mislabel, two release dates, and the name collision with the Seedance video line. Each claim, addressed in turn.
The official 100-round demo covers a single portrait; the blog's two other multi-turn demos (floral, interior) carry no round count, and no quality/fidelity curve over rounds has been published. The longest independent test is ifanr's 18 rounds; Seedance.io's 6-round drift test shows 5.5 much steadier than 5.0 Pro but not frozen — the label paper still darkened slightly after 6 rounds. "100 rounds" is a vendor claim, not a reproduced fact.
The API spec caps total pixels at about 4.62MP, 2K maximum; the 4K figures (~8.29MP) on OpenArt, Morphic and similar pages contradict the published API spec. The 3K tier some pages attribute to WaveSpeed conflicts with the API spec too — this site doesn't list 3K.
November 12 is OpenRouter's listing date (its page says "released November 12"); the model launched November 9 (Beijing afternoon, official blog + X announcement).
Only on editing: Seedance.io measured single-reference edit 2K at 61s median for 5.5 vs 97s for 5.0 Pro (about 38% faster). On text-to-image 5.5 is slightly slower than 5.0 Pro (47s vs 43s median). A blanket "40% faster" is wrong.
Seedance is ByteDance's video line; the latest is Seedance 2.5 from July 31, and no new Seedance launched on November 9. "Seedance 5.5" in Google Trends is a mistaken search.
The official blog only says "A Pro version of the Seedream 5.5 series will follow" — no date, no specs. Don't mistake 5.0 Pro (released July 8) for 5.5 Pro.
The GPT Image 2.5 Sunburst named in the blog is indeed the older generation (its successor 2.6 Sunburst joined Arena Nov 4); but the blog says "leading models such as GPT Image 2.5 Sunburst and Nano Banana 2.1" — it never says the comparison was limited to those two, so whether 2.6 was included can't be confirmed, and there's no basis to say ByteDance deliberately picked the older generation. The vendor's claimed lead holds only under the disclosed comparison conditions; comparison against the newest generation is unverified.
5.5 has no public weights — it's a hosted-only model that cannot be self-hosted or audited. "Available via API" is not "open source".
November 9, 2026 (Beijing afternoon): the Seed official blog + X announcement, with Dreamina, Doubao, Doubao Work and the Volcano Ark experience center going live the same day. The November 12 date on OpenRouter's page is its listing date, not the model release. See section 02 for the timeline.
China: Volcano Ark is live with model ID Doubao-Seedream-5.5, price list public since Nov 12. International: BytePlus's ModelArk model ID and per-image price are still unconfirmed as of Nov 14 (nowrap reports Dola-Seedream-5.5 listed, no price). Third-party APIs cover everything already: WaveSpeed (bytedance/seedream-v5.5), OpenRouter (bytedance-seed/seedream-5-5), Ofox (volcengine/doubao-seedream-5.5) and more. See our API access status article.
Volcano Ark official prices: output ≤2.61MP ¥0.13/image, 2K ¥0.23/image, input image ¥0.05 each, text free; a 2K edit with one input image is about ¥0.28. Third-party resellers: WaveSpeed/OpenRouter $0.02 (1K/1.5K) / $0.035 (2K) + $0.008 per input image; SeedRouter $0.017/$0.02975. BytePlus sells only image packs (700 for $6.99 first purchase), no per-image price. See our pricing guide.
The vendor claim is real (the blog states 100 rounds), but the evidence is a single portrait demo; no quality/fidelity curve over rounds has been published. The longest independent test is ifanr's 18 rounds; Seedance.io's 6-round drift test shows 5.5 much steadier than 5.0 Pro but not frozen. 100 rounds has not been independently reproduced. See our 100-round editing fact-check.
Yes: native RGBA, direct output of transparent PNGs with an alpha channel, fully and semi-transparent, generatable from text alone (5.0 Pro could only preserve existing transparency). Background modes are auto / transparent / opaque; transparent output is always PNG. But three independent tests agree: glass, water and other see-through objects themselves come out essentially opaque — clean cutouts, opaque glass. See our transparent PNG article.
Same 1K/1.5K/2K tiers, same 10-reference-image cap, same one image per request; 5.5 adds transparent PNG generation from text, steadier multi-round editing, more accurate small text, and layer decomposition from 16 to ~100 layers; output prices ¥0.13/¥0.23 vs ¥0.30/¥0.60 (2K about 62% cheaper). But 5.5 charges for every input image (5.0 Pro's first was free, then ¥0.02), so multi-reference edits narrow the gap. AA scores: text-to-image 1135 vs 1102, editing 1159 vs 1129.
On AA scores GPT Image 2.6 Sunburst is higher (text-to-image 1261, editing 1209–1210, vs 5.5's 1135/1159); on features 5.5 has 100-round editing, native RGBA and layer decomposition, while 2.6 has native GIF, SVG export and markup-layer editing. MaverickFrame and BeatAPI agree: long feedback chains + transparent assets + dense Chinese text → 5.5; GIF/SVG/precise point-and-edit → 2.6. See our comparison article and the GPT Image 2.6 topic page.
On AA text-to-image Nano Banana 2.1 is higher (1176 vs 1135); on editing they're statistically tied (1152 vs 1159, overlapping confidence intervals). ByteDance claims a lead over it on text structure (vendor evaluation, not independently confirmed). Quantslant's Polish short-headline test: Nano Banana 2.1 was 30/30 error-free, 5.5 was 27/30.
Jimeng web: Create → Generate/Canvas → Seedream 5.5; the launch promo (unlimited 1.5K for 7 days non-members / 30 days members) is single-sourced and may have ended. International Dreamina: the model page is live but access rolls out account by account. Doubao app/web: image generation → Seedream 5.5.
No official free tier could be confirmed: Hokai found no free tier on any official page. Dreamina's launch unlimited-use promo may have ended. SeedVideo AI gives new accounts free credits usable on 5.5 (a reseller perk, not an official free tier). BytePlus's Starter pack — 700 images for $6.99 first purchase, ~$0.01/image — is currently the cheapest official channel.
2K maximum, no 4K: 1K/1.5K/2K tiers, 11 aspect ratios, custom sizes capped at about 4.62MP total; the most pixels come from 16:9 at 2816×1584. The 4K figures (~8.29MP) on some pages contradict the published API spec. Note that 5.0 Lite actually has 2K/3K/4K tiers and up to 14 reference images — at the cost of the worst small text.
None: Seedream is the image line, Seedance is the video line; no new Seedance launched on November 9 — the latest Seedance is 2.5 from July 31. "Seedance 5.5" is a mistaken search.
All facts on this page come from the public sources below (checked Nov 14, 2026).