Doubao Explained: ByteDance’s AI Assistant
Doubao is the most-used AI app in China, and it is also the one most readers outside China cannot use at all. That contradiction is the whole story of this guide. Doubao (豆包, literally “bean bun”) is ByteDance’s AI assistant, and as of September 2026 it had 419 million monthly active users, according to QuestMobile — more than any other AI app in China, and more than most AI apps anywhere. Yet the app itself is published only in mainland China, Hong Kong and Macau. You cannot download it in the United States, and you cannot download it in most of Europe either. What reaches the rest of the world is not the app but the models underneath it, sold under different names — Seed, Seedance, Seedream — through BytePlus ModelArk, and a separate overseas assistant called Dola that runs on the same technology.
This guide is not a rewrite of our earlier introduction to Doubao. That post, from 18 October 2026, covered what the app is and how it became China’s most popular AI assistant. What it could not cover — because none of it existed yet in October — is the subject of this guide: the full Doubao-Seed model family, generation by generation, the Seedance and Seedream creative models with their November 2026 launches, the paid Doubao Pro tiers, and the precise map of where ByteDance’s models are and are not available outside China. If you want the short version of what the app does, start with our earlier explainer of what Doubao is; this guide is the model-family and availability explainer that sits on top of it. Checked 18 November 2026.
Three confusions need clearing up before anything else, because they trip up almost every English-language account of Doubao. First, “Doubao” is three things at once: the consumer app, the model family behind it, and the brand under which ByteDance’s image and video models reach consumers. Second, “Seed” is both the research team that builds those models and the international brand for the same models — the word “Doubao” appears nowhere in BytePlus’s international catalogue. Third, Doubao 2.0 is not the latest model. The current flagship is Doubao-Seed-2.1, released in June 2026, and the 2.0 line retires on 24 November 2026. Keep those three distinctions straight and the rest of this guide falls into place.

What Doubao Actually Is — One Name, Three Different Things
<img src="

” alt=”One object wearing three different hats, a visual metaphor for one name meaning three things: an app, a model family, and a brand. No text.”>
Doubao is ByteDance’s general-purpose AI assistant, launched in August 2023. As an app, it does what the leading assistants do: text chat with web search, writing and editing, homework help, translation, document analysis, slides, image and video generation, voice and video calls, and office tasks. It runs on iOS, Android, web and desktop, and its legal operator is Beijing Chuntian Zhiyun Technology Co., Ltd. By November 2024 it was already China’s most popular AI chatbot, with around 60 million monthly active users, according to industry analytics — and it has grown roughly sevenfold since.
The second meaning is the model family. The models behind the app are built by ByteDance Seed, a research team founded in 2023 with labs in China, Singapore and the United States, and they power more than 50 ByteDance apps, including Doubao itself, Coze and Jimeng. The family is called Doubao-Seed in China: Doubao-Seed-2.1 Pro is the current flagship text model, and the creative siblings are Seedance for video and Seedream for images. The same models are sold internationally under the Seed names, without the Doubao word attached — a branding split that matters, because a developer looking for “Doubao” in BytePlus’s international catalogue will not find it.
The third meaning is the consumer brand. When ByteDance’s image and video models reach ordinary users, they do so as Doubao features: Seedream makes the pictures inside the Doubao app, Seedance makes the videos. So “Doubao” is at once the product you open, the engine inside it, and the name on the box. The practical consequence is that almost every question about Doubao needs a second question — which Doubao? “Can I use Doubao?” has a different answer for the app (only in China, Hong Kong and Macau) than for the models (yes, through BytePlus, under Seed names, in more than 185 countries). “Is Doubao open source?” has a different answer for the flagship models (no, they are closed and API-only) than for the one open exception, Seed-OSS-36B (yes, Apache 2.0). This guide keeps the three meanings apart throughout, and says which one it means whenever the answer depends on it.
419 Million Users: The Scale Nobody Outside China Sees

The headline number comes from QuestMobile’s China Mobile Internet Autumn Report, released on 3 November 2026: Doubao had 419 million monthly active users in September 2026. Its users averaged 72.0 uses and 132.5 minutes per user per month, down 6.1% and 7.8% respectively on the previous quarter. Native AI apps in China as a whole reached 546 million monthly active users, up 82.1% year on year. These are QuestMobile measurements, not company disclosure — ByteDance has not confirmed the figures — and this guide attributes them as such. An earlier QuestMobile point puts the growth in perspective: 382 million monthly active users in June 2026, up 172.1% year on year, with around 155 million weekly active users and a peak near 145 million daily active users during the 2026 Lunar New Year. Qwen was the second AI app in China at around 167 million monthly active users in June 2026.
There is a twist in the September numbers, and it is the freshest part of the story. Doubao is still China’s number one AI app — but it is no longer ByteDance’s number two app overall. Hongguo, ByteDance’s free short-drama app, reached 430 million monthly active users in September 2026 and passed Doubao, sitting behind only Douyin at 1.029 billion. Our report on Hongguo passing Doubao gives the day-one detail; the point for this guide is that the “most-used” framing needs qualifying. Doubao is the most-used AI app in China. It is no longer the second most-used ByteDance app of any kind.
What does that scale buy ByteDance? The honest answer is a feedback loop and a distribution channel. Hundreds of millions of users, using the app more than twice a day on average, generate the usage data and the habits that shape the next model generation — and they give ByteDance a place to put every new capability the moment it is ready, from video generation to office agents to travel booking, without needing anyone to install anything new. That is the structural advantage a rival without a giant consumer app cannot copy by building a better model alone. It is also why the model family matters as much as the app: the same models that serve 419 million people in the Doubao app are the models ByteDance sells to developers worldwide, under different names, through BytePlus.
How Doubao Got There: Distribution, Free, and Fast Iteration

Doubao did not grow to 419 million users on model quality alone. It grew because ByteDance already owned the distribution. Douyin — TikTok’s Chinese sibling, with 1.029 billion monthly active users — is the largest funnel in Chinese consumer technology, and Doubao sits inside that ecosystem rather than beside it. Our earlier account of how Doubao became China’s number one AI app covers the distribution story in full, and this section does not repeat it; the short version is that ByteDance seeded an assistant inside the apps hundreds of millions of people already opened every day, then let the assistant earn its place. How Doubao became China’s number one AI app.
The second ingredient was that Doubao was free, and free in a way that mattered to mainstream users rather than to enthusiasts. For nearly three years the app charged nothing at all: search, writing, image generation and voice and video conversation were all free, with no subscription tier in sight. That changed on 24 June 2026, when Doubao Pro launched — the paid tiers are covered in their own section below — but the free core remains, and the three free years did the work of habit-building before any payment was asked for. The third ingredient was voice. Doubao was built voice-first for a mainstream audience that includes people who do not type long prompts, and its full-duplex voice calling and simultaneous-interpretation features made it usable in situations — a phone call, a car, a homework session — where a text-only assistant would not have been.
The fourth ingredient is iteration speed, and it is the one that shows up most clearly in the model timeline in the sections below: five flagship text generations in under two years, a video line that went from first release to a 30-second single-pass model in about thirteen months, and an image line that went from separate generation and editing models to a unified model and then to 100-round editing in about fourteen months. ByteDance’s research arm, Seed, ships at a pace that makes any single release a snapshot rather than a settled fact — which is why every date and version in this guide is stamped as checked on 18 November 2026, and why the guide names the retirements as well as the launches. Distribution got Doubao its users, being free kept them, voice made it mainstream, and iteration speed is what keeps the product moving faster than its rivals can copy it.
Doubao is also moving into the car. It works as an in-car voice assistant in mainland-China Tesla offerings, and ByteDance’s offline AI box with Pateo, running on NVIDIA Thor, brings Doubao into vehicles without a network connection. Our report on Doubao’s offline AI box for cars covers that in detail. The car is a small part of the story today, but it shows the pattern: wherever ByteDance can place an assistant, it places one.
Inside the App: What Doubao Does Day to Day

Open the app and the first thing you meet is a chat box with web search built in, and that is still the centre of the product: ask a question, get an answer grounded in the live web, follow up in conversation. Around it sit the features that make Doubao a daily tool rather than a novelty. Writing and editing, including long-form drafting. Homework help — photograph a problem and get worked solutions. Translation, including simultaneous interpretation on calls, with one mode that keeps the speaker’s own voice. Document analysis and slide and deck creation. Image generation through Seedream and video generation through Seedance, both inside the same app. Voice and video calls, with full-duplex voice, and a real-time video calling upgrade through SeedRealtime that arrived on 6 August 2026, free for all users. Office work through Feishu integration: read meeting outcomes and group chats, draft weekly reports. Travel advice and hotel booking. And on the PC client, an agent-driven office task mode that, after explicit user authorisation, drives local apps, the browser and files to produce Feishu documents, spreadsheets and presentations.
Two changes in 2026 reshaped what the app is allowed to do. On 15 July 2026 — the day China’s Interim Measures for the Administration of Anthropomorphic AI Interactive Services took effect — Doubao stopped allowing users to create custom agents, pointing them instead to ByteDance’s separate Maoxiang app; third-party agents remain available. And on 30 July 2026, Feishu’s product team was merged into Doubao, with the enterprise edition of Doubao running inside Feishu. The direction of both changes is the same: Doubao is becoming the single assistant for ByteDance’s consumer and office products, while the more open-ended agent experimentation moves elsewhere.
The limits are as important as the features, and they are limits of language and geography rather than of capability. Doubao is Chinese-language-first: it is built for Chinese users, in Chinese, with Chinese services behind it. Sign-up effectively needs a Chinese phone number — login is by mobile number, carrier one-click, Apple ID, or Douyin account — and user data is stored in mainland China, according to the privacy policy. The travel features come with a caution worth stating plainly: we put Doubao’s hotel booking to the test in October 2026, and readers should read that report rather than take the app’s travel advice at face value. Our test of Doubao’s hotel booking and travel advice. For a reader outside China, the honest summary of this section is that the app is a remarkably complete assistant — and one you almost certainly cannot download.
The Model Underneath: Doubao-Seed, Generation by Generation

The models behind Doubao are built by ByteDance Seed, founded in 2023, and they power more than 50 ByteDance apps. The flagship text line has moved through five generations in under two years, and the table below is the whole story in one place. Two generations deserve a note before the table: Doubao 1.8, announced on 18 December 2025 at Volcengine’s winter FORCE conference by president Tan Dai, is the generation most English-language coverage omits, and it brought agent capability and multimodal understanding; and Doubao-Seed-2.0, released on 14 February 2026, was described in Chinese coverage as the largest upgrade in the 21 months since the Doubao model launched.
| Generation | Date | Headline | Status, checked 18 Nov 2026 |
|---|---|---|---|
| Doubao-1.5-pro | 22 Jan 2025 | Sparse mixture-of-experts flagship; thinking variants 17 Apr 2025 | Superseded |
| Doubao 1.6 | 11 Jun 2025 | 1.6, 1.6-thinking and 1.6-flash; claimed 63% cost cut | Flash and Vision retired 21 Sep 2026 |
| Doubao 1.8 | 18 Dec 2025 | Agent capability and multimodal understanding; announced at Volcengine winter FORCE | Superseded |
| Doubao-Seed-2.0 | 14 Feb 2026 | Pro, Lite, Mini and 2.0-Code; largest upgrade in 21 months, per Chinese coverage | Pro and Code deprecated; retire 24 Nov 2026 |
| Doubao-Seed-2.1 | 23 Jun 2026 | Pro and Turbo at Volcengine summer FORCE; current flagship | Current |
The retirements matter as much as the releases. Doubao-Seed-1.6 Flash and 1.6 Vision were retired on 21 September 2026, with new purchases stopped on 10 July 2026; the migration targets are 2.0 Mini and 2.0 Lite. Doubao-Seed-2.0 Pro and 2.0 Code are deprecated and retire on 24 November 2026 — six days after this guide was checked — with migration targets of 2.1 Lite and Seed Evolving. OpenRouter’s bytedance-seed/seed-1.6 was retired on 11 November 2026. Anyone building on a Doubao model needs to read this table as a warning as much as a history: ByteDance retires generations quickly, and a model that works today may be gone within months. The migration targets are the point — ByteDance expects users to move forward, not to stay put.
The scale of usage behind the family is, by ByteDance’s own account, enormous. Volcengine president Tan Dai said at the June 2026 FORCE conference that daily token calls across the Doubao model family had reached 180 trillion tokens per day, more than ten times a year earlier. IDC put Volcengine at 49.5% of China’s enterprise model-as-a-service market for full-year 2025 by public-cloud token call volume — first in the country, ahead of Alibaba Cloud and Baidu AI Cloud, up from 46.4% in 2024. Both figures are vendor and analyst claims, and this guide attributes them as such rather than presenting them as independent measurement. But even discounted, they describe a model family operating at a scale no Western provider’s enterprise business matches in its home market — and they explain why the pricing in the next section is as aggressive as it is.
Doubao-Seed-2.1: The Current Flagship, Explained

Doubao-Seed-2.1, released on 23 June 2026 at Volcengine’s summer FORCE conference, is the current flagship generation, and it comes in three tiers. Doubao-Seed-2.1 Pro, model ID doubao-seed-2-1-pro-260628, launched with a 256,000-token context window and was refreshed to a 1-million-token context on 15 September 2026, under model ID doubao-seed-2-1-pro-260915, at the same price: CNY 6 per million input tokens and CNY 30 per million output tokens on Volcengine Ark. Maximum output is 256,000 tokens, with a 4,000-token default, and maximum chain-of-thought is 256,000 tokens. Doubao-Seed-2.1 Turbo, model ID doubao-seed-2-1-turbo-260628, is the half-price tier at CNY 3 per million input tokens and CNY 15 per million output tokens, with a 256,000-token context. Cache hits are 80% cheaper than full-price input across the line.
The third tier arrived with the September refresh. Doubao-Seed-2.1 Lite, released on 15 September 2026 as doubao-seed-2-1-lite-260915, is omni-modal — text, vision and audio in a single request — with a 1-million-token context, 256,000-token maximum output, and a price of CNY 0.80 per million input tokens and CNY 2.70 per million output tokens, with cache hits at CNY 0.16. It is integrated into Doubao Work for office tasks, and it is the migration target for users leaving the retiring 2.0 Pro. The pattern across the three tiers is deliberate: Pro for the hardest work, Turbo for the same capability at half the price, Lite for the broadest, cheapest, most multimodal use. Prices are in CNY on Volcengine Ark, the Chinese platform; the international prices through BytePlus ModelArk are listed under Seed names and are covered in the availability sections below.
Two more members of the family need explaining, and one needs a warning. Doubao-Seed-Evolving is a single stable model ID that is updated at least weekly, with no version numbers, priced like the flagship at CNY 6 per million input tokens and CNY 30 per million output tokens — but its context window is disputed between ByteDance’s own surfaces. Volcengine’s model list says 1,024,000 tokens; its AI Hub catalogue says 256,000 tokens. This guide reports both figures and picks neither, because the disagreement is on ByteDance’s own pages and no outside source can resolve it. Doubao-Seed-Code, released on 11 November 2025, is the dedicated coding model: 256,000-token context, Anthropic-API compatible, priced at CNY 1.20 to 2.80 per million input tokens and CNY 8.00 to 16.00 per million output tokens, tiered by prompt length. Doubao-Seed-Translation, released on 22 September 2025, covers 28 languages at CNY 1.20 per million input characters. The coding and translation models get their own treatment in the family section below; the point here is that 2.1 is not one model but a tiered product line, and the tier you choose changes the price by an order of magnitude.
A word on benchmarks, because this is where most guides about Doubao go wrong. All benchmark figures for Doubao-Seed models are vendor-reported, and none has been independently verified. Doubao-Seed-2.0 Pro was announced with a score of 54.2 on the text subset of Humanity’s Last Exam, which ByteDance characterised at its February 2026 launch as the leading score globally — that is ByteDance’s characterisation, not an independent ranking, and this guide states it as such. For 2.1 Pro, Volcengine claimed parity with a leading Western frontier model on Terminal Bench 2.1, plus scores of 59.8 on SciCode and 47.0 on NL2Repo-Bench, and claimed leads on MCP-Atlas; Xinhua’s account of the launch carried no numerical scores at all. And as of 23 August 2026, no Doubao or Seed model appears on the public arena.ai text leaderboard — Kimi, ChatGPT, Qwen, Grok and DeepSeek all place; Doubao is absent. This guide therefore prints no benchmark league table and states no head-to-head verdict against ChatGPT, Gemini, Claude or DeepSeek on scores. The comparison in the later section is on access, language, ecosystem, price and openness — the dimensions where the evidence is real.
Seedance: The Video Models, From 1.0 to 2.5 Flash

Seedance is ByteDance Seed’s video-generation line, and it is the fastest-moving part of the whole family. The technical report appeared on arXiv on 10 June 2025, and Seedance 1.0 followed on 11 June 2025: text-to-video and image-to-video, multi-shot, 1080p output, with the report claiming a five-second 1080p clip in 41.4 seconds on an NVIDIA L20. On BytePlus ModelArk it appeared as Seedance 1.0 Lite in June 2025 and 1.0 Pro in July 2025, at 480p, 720p and 1080p at 24 frames per second, for clips of 2 to 12 seconds. That was the foundation; every release since has added a capability that the previous one lacked, rather than merely improving the picture.
Seedance 1.5 Pro, released on 16 December 2025, added native joint audio-video generation — sound generated with the picture, not added afterwards — with multilingual lip-sync. Seedance 2.0, released on 12 February 2026, went further: a unified multimodal audio-video model taking text, image, audio and video as inputs, integrated across the Doubao app, desktop and web on the same day, with 4K output and IP guardrails and C2PA watermarking. Then came the current flagship. Seedance 2.5, launched on 31 July 2026 after a preview at the 23 June FORCE conference, generates 30 seconds in a single pass with sound and multiple connected shots, with region-level editing and deeper reference handling. It is the model that turned Seedance from a clip generator into something closer to a scene generator.
The freshest release is the one this guide leads with. Seedance 2.5 Flash launched on 12 November 2026 — six days before this guide was checked — as an accelerated cut of 2.5: clips of 4 to 30 seconds at 480p or 720p with sound, with no 1080p or 4K option, up to 20 references (10 images, 5 video, 5 audio, each capped at 30 seconds), at roughly twice the speed of the flagship and roughly $0.15 per second of 720p video, about 34% below the flagship’s rate, according to BytePlus ModelArk pricing. It is live in Dreamina’s AI Video, AI Canvas and AI Agent workflows, rolling out to Jimeng and others, and available through the API via BytePlus ModelArk (listed 15 November) and Volcengine Ark under model ID doubao-seedance-2-5-flash-261115. Our report on Seedance 2.5 Flash gives the launch detail. One figure this guide deliberately does not print is the flagship Seedance 2.5 per-second price: only the Flash tier price is on record in the sources for this guide, and the flagship rate is not stated.
Two wider facts complete the picture. Seedance’s overseas revenue share reached 50%, according to 36Kr reporting in November 2026 — a single-source figure, reported here as reported, not confirmed — which would make the video line ByteDance’s most international AI product by revenue. And in August 2026 ByteDance and the Motion Picture Association signed a memorandum of understanding on IP protection for Seedance and Seedream, a sign that the creative models are now big enough to need formal agreements with the industries they disrupt. For readers comparing video generators more broadly, our comparison of AI video generators sets out how Seedance sits against the rest of the field.
Seedream: The Image Models, From 4.0 to 5.5

Seedream is the image line, built by the Seed Vision team, and its story is one of unification followed by endurance. Seedream 3.0, released in April 2025, generated images, and SeedEdit 3.0 edited them — two separate models. Seedream 4.0, released on 9 September 2025, unified them into one model: text-to-image generation, instruction-based editing and multi-image composition in a single model, with native 4K output, up to 10 reference images, and a reported 12-billion-parameter mixture-of-experts design. The vendor claim was a 2K image in under two seconds. The price most often quoted, around $0.03 per image through BytePlus, is a reseller and aggregator rate, not a ByteDance-published USD price, and this guide says so.
The version confirmed inside the consumer app is Seedream 4.5, released on 3 December 2025 by Volcengine and made available inside the Doubao app, Jimeng and Volcano Ark, with improved multi-image editing accuracy. That distinction — confirmed in-app versus merely announced — is the one this guide is careful about. Seedream 5.0 Lite arrived in February 2026 with a reasoning and intent focus, and Seedream 5.0 Pro was announced on 23 June 2026 at the FORCE conference, with its API opening on 8 July 2026, framed as an enterprise and developer product. A Seedream 5.0 Flash was listed on 29 September 2026 in an aggregator catalogue, a secondary source. But no confirmation exists that 5.0 Pro has shipped inside the consumer Doubao app: the version confirmed in-app is 4.5, and this guide does not state a current in-app version beyond that.
The freshest release is Seedream 5.5, launched on 9 November 2026 — nine days before this guide was checked — under the ByteDance Seed blog headline “More Realistic Images, More Stable Editing Across 100 Rounds”. Its headline capability is stable multi-turn editing: up to 100 consecutive edits with unedited regions preserved at pixel level. It also produces native RGBA transparent output and denser small text, with a vendor-reported severe text-corruption rate of 0.4% and a minor rate of 1.7%. It is rolling out in Jimeng, Doubao, Doubao Work and Volcengine Ark, with a Pro version promised but undated. No published model ID or official USD price existed as of 9 November 2026, and this guide states neither. Our report on Seedream 5.5 gives the launch detail. The honest way to describe the image line today is that 5.5 is rolling out in the app, 4.5 is the last version confirmed in-app, and 5.0 Pro is a developer product — three statements that are all true at once, and that no single version number can summarise.
The Rest of the Family: Code, Audio, Translation — and the One Open Model

The specialist siblings are easy to miss beside the flagship text models and the creative lines, but they are where the family shows its range. Doubao-Seed-Code, released on 11 November 2025, is the dedicated coding model: a 256,000-token context, Anthropic-API compatibility, and tiered pricing of CNY 1.20 to 2.80 per million input tokens and CNY 8.00 to 16.00 per million output tokens, depending on prompt length. It is the model behind the coding assistance inside Doubao and its developer products, and its Anthropic-API compatibility is a deliberate choice — code written against Anthropic’s API can be pointed at it with a changed endpoint.
Seed-Audio 1.0, announced at the June 2026 FORCE conference, is an end-to-end model generating up to two minutes of audio per call: dialogue, emotion, music, ambient sound, and multi-speaker timbre consistency. It remains in invite-only API beta, which is worth stating plainly — it is not a product you can sign up for today. Doubao-Seed-Translation, released on 22 September 2025, covers 28 languages at CNY 1.20 per million input characters, and it is the engine behind the simultaneous interpretation on calls described in the app section. Together, the three specialists cover the three jobs the flagship text model does not do best: writing code, making sound, and crossing languages.
Then there is the exception that proves the closed rule. Seed-OSS-36B-Instruct, released on 20 August 2025, is the one open model in the family: 36 billion parameters, a native 512,000-token context, released under Apache 2.0 on Hugging Face under the ByteDance-Seed organisation. Its model card reports 82.7 on MMLU-Pro, 91.7 on AIME24, 94.6 on RULER at 128,000 tokens, and 56 on SWE-bench Verified under OpenHands — self-reported but reproducible figures, stated here as reported. Further research checkpoints followed, including Cola-DLM, described in a paper dated 7 May 2026. The rule to remember is simple and absolute: the flagship Doubao models are closed and API-only, and no “Doubao open-source model” exists. Seed-OSS-36B is open; everything with “Doubao-Seed” in its name at the top of the range is not. And one more thing that is not a product: Seed-Music is a research framework, described in a paper from September 2024, not a product — do not look for it in any catalogue.
Free No More: Doubao Pro and What Paying Actually Buys

Doubao was entirely free for nearly three years. That ended on 24 June 2026, when Doubao Pro (豆包专业版) launched in mainland China only — the app’s first paid consumer subscription. The everyday functions stay free: search, writing, image generation, voice and video conversation all remain available without paying. What paying buys is quota, plus the PC office task mode — the agent-driven feature that drives local apps, browser and files to produce Feishu documents, spreadsheets and presentations. The free tier is the product most people know; the paid tiers are for the people who use it hardest, and for office work.
The tiers, as listed in the mainland App Store and reported at launch, are auto-renewing subscriptions in CNY. Standard (标准套餐) costs CNY 68 per month, or CNY 688 per year, and offers at least five times the free allowance, plus access to the Doubao 2.1 Pro model. Enhanced (加强套餐) costs CNY 200 per month, or CNY 2,048 per year, at roughly four times Standard’s allowance. Premium (高级套餐) costs CNY 500 per month, or CNY 5,088 per year, at roughly ten times Standard’s allowance. A student rate of CNY 38 per month was offered for verified mainland university students, as a six-month promotion announced at launch. A 30-day trial was offered on the desktop download page as of 5 October 2026, but that was not verified in this session and is not stated here as current. No official public price list exists on doubao.com, so these prices are as listed in the mainland App Store and reported at launch, and this guide phrases them that way. A conversion put the monthly tiers at roughly $9.40, $27.60 and $69 — that conversion is attributed to TechNode, and no other USD figure is given, because converting CNY to USD with an unattributed rate would be inventing a price.
How does that compare with what Western assistants charge? The honest answer is that the comparison is on structure, not on price alone. ChatGPT, Gemini and Claude all charge a flat monthly subscription for their paid tiers, with the model included; Doubao Pro charges for quota on top of a free product, with the top tier aimed squarely at office work rather than at chat. For the Western paid tiers in detail, our comparison of ChatGPT’s free, Plus and Pro tiers is the closest parallel. The point for a reader outside China is narrower: Doubao Pro is mainland-China only, like the app itself, so the tiers are a fact about the Chinese market, not an offer you can take up. What matters is the direction — ByteDance has decided that its most-used consumer AI product is worth charging for, and that the thing worth charging for is not chatting but working.
China vs the World: Where You Can (and Can’t) Use Doubao

The Doubao app is published to exactly three Apple App Store storefronts: mainland China, Hong Kong and Macau. It launched in mainland China in August 2023 and reached Hong Kong and Macau in November 2024. The Android package, com.larus.nova, resolves in some storefronts — the UK and Singapore are confirmed — but as a Chinese-language listing. Sign-up effectively needs a Chinese phone number, and data is stored in mainland China under the privacy policy. And doubao.com, visited from outside China, shows a regional-restriction notice and points visitors to Dola. That is the whole consumer story in one paragraph: the world’s most-used AI app is, as a consumer product, unreachable in the United States and in most of the world.
Dola is ByteDance’s overseas assistant, and it is a different app, not a translation of Doubao. It launched in August 2023 as Cici and was renamed to Dola around November 2025. It has an English interface and around 18 languages, and it is available in the UK, Singapore, Malaysia, Mexico, Indonesia, Philippines, Brazil, Hong Kong, Thailand, Spain, France, Poland, the UAE, Saudi Arabia, Egypt and Argentina, among others — but not in the United States, Canada, Australia, India or mainland China. It shares a version number and much of a codebase with Doubao, but some Doubao features, such as native audio-synced video generation, are blocked in it. Until February 2026 Dola ran on OpenAI GPT and Google Gemini models; it switched to in-house Seed models from February 2026, rolling out fully by April 2026. It was reported to have more than 10 million daily active users worldwide by the end of 2025 — a secondary, single-thread figure, stated here as reportedly.
For developers, the map is different again. BytePlus ModelArk sells the same models under Seed names — seed-2-0-pro, dola-seed-2-1-turbo-260628, seedream-*, Seedance-1.5-pro — and the word “Doubao” appears nowhere in the international catalogue, as checked on 23 August 2026. ModelArk documents availability in more than 185 countries, excluding the United States, mainland China and Russia, with regions in Johor (ap-southeast-1) and Dublin (eu-west-1). But the international catalogue trails the Chinese one: as of August 2026 it listed no seed-2-1-pro and no equivalent of doubao-seed-evolving. Six Seed models are also available on OpenRouter under bytedance-seed/. The summary is sharp: as a consumer product, Doubao stops at China’s borders; as a developer product, its models travel the world under a different name, with a lag, and with the United States excluded.
How to Try ByteDance’s Models From Outside China

There are three ways in from outside China, and the right one depends on what you want: a consumer assistant, an API, or a model on your own machine. Start with the easiest and move down only when it stops being enough. The first route is Dola, the overseas assistant. If you are in one of its markets — the UK, Singapore, Malaysia, Mexico, Indonesia, the Philippines, Brazil, Hong Kong, Thailand, Spain, France, Poland, the UAE, Saudi Arabia, Egypt or Argentina, among others — you can download Dola and use an assistant built on the same Seed models as Doubao, with an English interface and around 18 languages. What you cannot do is download it in the United States, Canada, Australia, India or mainland China, and some Doubao features, such as native audio-synced video generation, are blocked in it. Dola is the closest thing to Doubao that exists outside China, and it is not Doubao.
The second route is the API. BytePlus ModelArk sells the Seed models to developers in more than 185 countries, excluding the United States, mainland China and Russia, with regions in Johor and Dublin. Search the catalogue for Seed names, not Doubao — seed-2-0-pro, seedream-*, Seedance-1.5-pro — because the word “Doubao” appears nowhere in it. Expect the catalogue to trail the Chinese one: as of August 2026 there was no seed-2-1-pro and no equivalent of doubao-seed-evolving listed, so the newest flagship may not be the one you can call. Six Seed models are also listed on OpenRouter under bytedance-seed/, which is the simplest way to try them if you already use that platform. The third route is local: Seed-OSS-36B-Instruct, released on 20 August 2025 under Apache 2.0, is the only self-hostable route in the family. It is a 36-billion-parameter model with a native 512,000-token context, available on Hugging Face under the ByteDance-Seed organisation. It is not the flagship, and it is not Doubao-Seed-2.1 Pro — but it is the only model in the family you can download and run yourself.
What you cannot do, from outside China, is the obvious thing: download the Doubao app, sign up with a non-Chinese number, and use the flagship model as a consumer. The app is published only in mainland China, Hong Kong and Macau; sign-up effectively needs a Chinese phone number; and doubao.com redirects you to Dola. That is not a workaround waiting to be found. It is the product’s design. The three routes above are the real ones, and the honest guide is the one that says so plainly rather than hinting at a loophole.
Doubao vs ChatGPT, Gemini, DeepSeek and Qwen

Doubao’s rivals split into two groups, and the comparison is different for each. Against ChatGPT and Gemini, the consumer comparison, the difference is access, language, ecosystem and price — not capability scores, because no independent benchmark exists for any Doubao or Seed model, and this guide states no verdict on scores. ChatGPT and Gemini are global, English-first, and available in the United States; Doubao is China-first, Chinese-language-first, and unavailable in the United States at all. Doubao’s differentiator is ecosystem distribution — Douyin, Feishu, CapCut and Jimeng — and scale in China, not benchmark leadership. A reader choosing between them is, in practice, choosing between two markets: the assistant you can use depends on where you live, and for most readers outside China the choice is made for them.
Against DeepSeek and Qwen, the comparison is about openness, and it is the deepest split in the Chinese AI industry. Doubao owns consumers — 419 million monthly active users — while DeepSeek owns developers and strategists through open weights, with around 130 million monthly active users in China according to industry trackers cited in our earlier comparison. Our explainer of DeepSeek gives the detail. Doubao’s flagship models are closed and API-only; DeepSeek’s are open-weight. Qwen sits in between: the number two AI app in China at around 167 million monthly active users in June 2026, with a flagship line that is mostly open-weight under Apache 2.0 — although Qwen’s newest flagship, Qwen4-Max, released on 16 November 2026, is also API-only. Our explainer of the Qwen model family covers that in full. The pattern across all three is that the newest flagship is closed, wherever it comes from, and the open line sits one generation behind.
Against Kimi, from Moonshot AI, the comparison is a single paragraph, because Kimi is a separate subject for a later guide. Kimi is the long-context specialist; Doubao’s 2.1 line now offers a 1-million-token context too, so the context gap has closed on paper. But a context window on a specification sheet is not a capability in an app, and the two products serve different audiences — Kimi as a focused tool, Doubao as an everything assistant inside a giant ecosystem. The short version of the whole section is that Doubao wins on distribution and scale in China, loses on availability everywhere else, and cannot be ranked on capability at all, because the evidence to rank it does not exist independently. Any guide that declares a winner between Doubao and ChatGPT or Gemini on the basis of vendor-reported scores is declaring a verdict it has not earned.
Should You Care About Doubao? A Verdict — and What to Watch Next

There is no single answer to whether Doubao matters, because it matters to different people for different reasons. It matters to anyone watching China, because the most-used AI app in the world’s largest internet market is a fact about the world, not a curiosity — 419 million monthly active users, a paid tier launched in June 2026, and a model family that its maker sells to developers in more than 185 countries. It matters to video and image creators, because Seedance and Seedream are now serious tools: a 30-second single-pass video model with sound, and an image model that survives 100 rounds of editing without losing the parts you did not touch. And it matters to developers outside China, because the Seed models on BytePlus ModelArk are the practical way to use ByteDance’s technology without living in its home market. For a wider framework on choosing between models for a given job, our guide to choosing an AI model sets out what speed, cost, reasoning depth and context actually trade off against each other.
| Who you are | What to use | Why |
|---|---|---|
| China-watcher | The app, in China, or the QuestMobile numbers | 419M MAU; the most-used AI app in China |
| Video creator | Seedance 2.5 / 2.5 Flash via Dreamina or ModelArk | 30-second single-pass video; Flash at ~$0.15/sec |
| Image creator | Seedream 5.5 via Jimeng, Doubao or ModelArk | 100-round editing; native transparency |
| Developer outside China | BytePlus ModelArk under Seed names | 185+ countries; no “Doubao” in the catalogue |
| Consumer outside China | Dola, where available | Same Seed models; not in the US |
| Self-hosting | Seed-OSS-36B-Instruct | The only open model in the family; Apache 2.0 |
The limits belong in the same breath as the recommendations. There is no Doubao app in the United States, and no Dola either. There is no independent benchmark for any Doubao or Seed model, so no capability ranking against ChatGPT, Gemini, Claude or DeepSeek can be honestly stated. The international catalogue trails the Chinese one, so the newest flagship may not be the one you can call. And the flagship models are closed, so the only model you can run yourself is a 36-billion-parameter open release from August 2025, not the current flagship.
What to watch next is short, and every item on it is unconfirmed. A Seedream 5.5 Pro version has been promised, with no date given. The next generation of the text line has no name and no date in this guide, and this guide predicts neither. The retirement of Doubao-Seed-2.0 Pro and 2.0 Code on 24 November 2026 is the nearest dated event, and it will move the remaining 2.0 users onto 2.1 Lite and Seed Evolving. The family moves every few months, and the check that matters is the one this guide performed on 18 November 2026: read the release notes, read the catalogue, and do not trust a guide — including this one — past its checked date. Checked 18 November 2026.
Sources
All sources below were read on 18 November 2026, except where noted. Facts, dates, prices and model details in this guide come from these pages. No competitor or comparison-site pages were used as sources. Prices checked 18 November 2026.
- doubao-english.com — The Doubao AI Model Family, Explained (2026; generations, prices, benchmarks, siblings, availability; last updated 28 Aug 2026; figures cross-checked against Nov 2026 pricing pages) — https://doubao-english.com/doubao-ai-models/
- doubao-english.com — Doubao Features and Capabilities (Aug 2026); Doubao Free vs Paid (24 Jun 2026 launch, tiers); Doubao Availability Checker (App Store storefronts, Dola markets); Doubao Image Generation: Seedream Models Explained (4.5 in-app, 5.0 Pro developer-only) — read via search-result full text, cross-checked across at least two sources each — https://doubao-english.com/
- Wikipedia — “Doubao” (launch Aug 2023; ~60M MAU Nov 2024; Pro tiers; agents discontinued Jul 2026; Dola ex-Cici; Tesla in-car) — read via search 18 Nov 2026 — https://en.wikipedia.org/wiki/Doubao
- QuestMobile China Mobile Internet Autumn Report (3 Nov 2026) — via China Daily (4 Nov 2026), Asia Daily, Longbridge, 36Kr, ChainCatcher — consistent on 430M / 419M / 1.029B, 72 uses, 132.5 min
- aicost.tools — Doubao API Pricing (Nov 2026, verified 29 Oct 2026); Doubao Seed 2.1 Lite / 1.6 Flash / 1.6 Vision pages (Volcengine catalogue reads; retirement dates)
- seed-video.com — What Is Seedance? (lineage 1.0 to 2.5, dates from the ByteDance Seed blog)
- ByteDance Seed blog — “More Realistic Images, More Stable Editing Across 100 Rounds” (Seedream 5.5, 9 Nov 2026); “Today, we are officially launching Seedance 2.5” (31 Jul 2026)
- Seedream 5.5 launch coverage (9–10 Nov 2026): rockit.site, aireiter.com, quantslant.com, magica.com, blog.sandbase.ai — consistent on 9 Nov date, 100-round claim, RGBA, 0.4% / 1.7%, no published model ID or price
- Seedream 4.0: jademond.com, unicornapi.net, llms.ninja — consistent on 9 Sep 2025, unified generation and editing, 4K
- BytePlus ModelArk / Volcengine Ark catalogue facts — via doubao-english.com, hop-base.com, api.haijingai.com — consistent on Seed naming, 185+ countries, 2.1 Lite 15 Sep 2026, 1M context
- GER news — “ByteDance’s Seedream 5.5 Survives 100 Edits Without Losing Itself” (10 Nov 2026) — https://geroundtable.com/seedream-5-5-100-round-edits/
- GER news — Seedance 2.5 Flash (15 Nov 2026) — https://geroundtable.com/bytedance-seedance-2-5-flash/
- GER news — “Hongguo Passes Doubao as ByteDance’s No. 2 App” (7 Nov 2026) — https://geroundtable.com/hongguo-overtakes-doubao-430-million-users/
- GER news — “Doubao Moves Into the Car With an Offline AI Box on NVIDIA Thor” (6 Nov 2026) — https://geroundtable.com/doubao-pateo-ai-box-nvidia-thor-car/
- GER news — “Doubao’s Hotel Fee Puts Its Travel Advice to the Test” (29 Oct 2026) — https://geroundtable.com/doubao-hotel-booking-fee-travel-advice/
- GER evergreen — “What Is Doubao? ByteDance’s AI Assistant With Hundreds Of Millions Of Users” (18 Oct 2026) — https://geroundtable.com/what-is-doubao-bytedances-ai-assistant/
- GER evergreen — “How Doubao Became China’s Number One AI App” (18 Oct 2026) — https://geroundtable.com/how-doubao-became-chinas-number-one-ai-app/
- GER evergreen — “Doubao Vs DeepSeek: ByteDance And The Open-Model Upstart” (18 Oct 2026) — https://geroundtable.com/doubao-vs-deepseek/


0 Comments