Doubao App 2.0

• 1. overview (621 chars)

  Doubao is a national-level AI smart assistant owned by ByteDance. The app, web page and desktop are interoperable, and its monthly active users have exceeded 300 million. In February 2026, the large model supporting Doubao was upgraded to 2.0 (Doubao-Seed-2.0), and its multi-modal understanding and complex task execution capabilities ranked first in the world. In June 2026, Doubao launched a professional version based on the 2.1 series, upgrading the "Conversation Assistant" to a "Productivity Agent" that can control local computers. This report understands "Doubao App 2.0" as the current form of Doubao App equipped with the 2.0/2.1 series of large models, covering its capabilities, pricing, reputation and controversy.

• 2. Product background (626 chars)

  Doubao was launched by Bytedance in 2024. It is the company's flagship consumer product at the AI ​​application layer. Together with Volcano Engine's Doubao large model API and AI programming product TRAE, it forms a full-link layout of "application-cloud-developer". As of December 2025, the average daily call volume of Doubao's large model exceeded 63 trillion Tokens, ranking first in China and third in the world. Volcano Engine's market share of domestic public cloud large model call volume reached 49.2%. At the Byte all-member meeting, CEO Liang Rubo set the company’s keyword for 2026 as “climbing the peak” and bluntly stated that the short-term peak is the “Doubao/Dola assistant application”, which shows its strategic weight.

• 3. Product functions and usage experience (1579 chars)

  The core capabilities of Doubao App revolve around "all-round life assistant". The text side supports writing, translation, long article summaries, and PPT generation; the multi-modal side supports image generation (image generation, reference image toning change), video generation and disassembly, photo recognition, voice cloning, and AI podcasting. Voice interaction is its most praised feature - the response delay is about 0.8 to 1.2 seconds, which is close to the rhythm of a real person. It supports more than 20 dialects such as Sichuan and Cantonese, and can accurately capture Internet terms such as "emo" and "YYDS". It can also directly parse the video links of Douyin and Xiaohongshu, summarize the scripts or generate brain maps, and deeply connect the byte system ecosystems such as Douyin and Toutiao to produce short video scripts and product copywriting with one click.   The key upgrade brought by the 2.0 large model is "complex task execution capability." Doubao 2.0 Pro won gold medals in IMO, CMO mathematics competitions and ICPC programming competitions. It led Gemini 3 Pro and GPT 5.2 with 54.2 points in the HLE (Human Final Examination) text list. Benchmarks such as visual reasoning, spatial perception, and long video understanding generally reached SOTA. Users can experience it within the app by selecting "Expert" mode.   The professional version launched in June 2026 takes the capabilities a big step forward. It is equipped with the 2.1 Pro model and has a new "Office Task Mode": it can operate the local computer, call browsers and Skills, and perform scheduled tasks. It has a built-in Office suite that can automatically generate PPT, produce short plays with one click, and organize invoices in batches. From "answering questions" to "doing things on behalf of others", this is Doubao's iconic transformation from assistant to agent.

• 4. Pricing and business model (1177 chars)

  The base of the bean bag is free for a long time. Basic functions - daily Q&A, short copywriting, basic Internet search, photo recognition, ordinary voice dialogue, low-definition pictures - are permanently free for everyone, and the official has made a clear commitment that they will not shrink with the launch of the professional version.   The professional version will be launched on June 24, 2026, with a three-level monthly subscription: standard continuous monthly subscription is 68 yuan (the quota for office tasks and expert mode is more than 5 times that of the free version), enhanced 200 yuan (the standard version is 4 times the quota), and advanced 500 yuan (the standard version is 10 times the quota, for enterprise developers and heavy office workers). After college students complete the real-name authentication, the standard version can enjoy an exclusive discount of 38 yuan/month for 6 months. The free version is based on the 2.1 Turbo model and shares computing power. It may queue up during peak periods and long texts are easily truncated; the professional version is based on 2.1 Pro, has exclusive priority computing power and responds in seconds throughout the day. The external API of the model layer also follows a low-price route: 2.0 Pro inputs 3.2 yuan/million tokens and outputs 16 yuan/million tokens, which has a clear cost advantage compared to overseas flagships.

• 5. User reviews (1256 chars)

  Positive comments focus on "speaking people's words and doing practical things." In the App Store and many Hengping reviews, users generally praise its natural Chinese interaction, good voice experience, and practical efficiency in office work. It is very easy to write weekly reports, take meeting minutes, tutor children, and dismantle Douyin video scripts. The interface is clean and ad-free. In the 2026 Hengping Review of Domestic AI Tools, Doubao scored 8.3 points and was rated as "the all-round AI with the best Chinese experience."   Negative feedback was directed at three areas. First, there is a gap between hard-core reasoning and code: extremely complex logic, mathematics competition questions, and large-scale system architecture codes are still not as good as DeepSeek, Claude Opus, or GitHub Copilot. Writing long articles of more than 800 words is easy to go off track. Second, cross-application operations will trigger strong risk control in other apps, so you need to be cautious in enterprise-level sensitive scenarios. Third, after the professional version was launched, the reputation reversed: the first batch of paying users complained that the desktop Agent frequently "dropped offline", the credit was consumed quickly and was opaque, it was interrupted when used up, and could not be handed over to other tools. Users of the 68 yuan and 688 yuan tiers felt "cheated", and refunds were blocked by the clause that "virtual goods are not refundable in principle."

• 6. Industry evaluation (1041 chars)

  The industry generally regards Doubao as a model for the "application + cloud" collaboration of domestic large-scale models. Securities Times, Science and Technology Daily and other media believe that Doubao 2.0, together with the Seedance 2.0 video model and Seedream 5.0 Lite image model released at the same time, have achieved full-stack breakthroughs in the three dimensions of vision, video, and images, and have been described as "shooting Sora and tough Nano Banana." Tan Dai, president of Volcano Engine, called the 2.1 Pro "the flagship model that has crossed a qualitative change point in productivity", and its coding and agent evaluation has stably exceeded Claude Opus 4.6. In the SuperCLUE evaluation, the bean bag Seed-2.0-Pro ranked first in the country with a score of 71.53.   In terms of competition, Doubao faces Tongyi Qianwen, Wenxinyiyan, Kimi, and DeepSeek in China. Its differentiation lies in its full functionality, deep ecosystem, and fast updates; Tongyi is strong in document parsing, Wenxin is strong in multi-modal recognition, Kimi is strong in long texts, and DeepSeek is strong in deep reasoning - each company has its own "optimal solution" scenario, and there is no one that can completely crush it.

• 7. Controversies and Risks (1018 chars)

  The controversy first came from the commercialization of the professional version. Paying users reported that the credit limit was not transparent, tasks were interrupted and tasks could not be relayed, and the platform’s format clauses preset “refunds are not supported in principle”. Legal experts pointed out that this may constitute an invalid “overlord clause” under the Civil Code and the Consumer Rights Protection Act, which poses regulatory and litigation risks.   The second is the integrity of mind at the boundary of ability. The promoted "Productivity Agent" will still overturn in lightweight image editing and stable long-running process tasks, which can easily cause paying users to feel a sense of value asymmetry and affect renewal.   Next is privacy and risk control. The office task mode requires in-depth reading of local files and browser manipulation. While cross-application operations trigger strong risk control, it also brings questions about data compliance and permission boundaries. How to ensure user privacy is a new question that the entire industry is still answering. Although Doubao has released a privacy white paper, its implementation is still in its early stages.

• 8. Applicable groups and usage suggestions (572 chars)

  Doubao is most suitable for parents of students, professionals and content creators: tutoring on shooting questions, writing copy, making PPT, dismantling short videos, and voice chatting. One App can replace several tools on your phone. For heavy office work, self-media mass production, and enterprise development, it is recommended to go to the professional version.   If your needs are to write code, do complex mathematical reasoning, or analyze very long professional documents, a safer choice is DeepSeek or Claude; the best way to use Beanbao is to "open Beanbao for daily use, and open DeepSeek for deep work." Both are free, so there is no need to choose one over the other.

• 9. Summarize (513 chars)

  Doubao App 2.0 is the representative of domestic AI assistants that retain users by "relying on full functionality, good experience, and fast updates." The 2.0/2.1 series of large models allow it to stand at the forefront of the world in multi-modal and complex tasks; although the professional version has started commercialization, the stability of the desktop Agent and the refund mechanism are still lessons it must make up for. If you treat it as an all-round life companion, it is very capable; if you treat it as a serious productivity main force, you still need to wait and see for another generation or two.

No in-depth report yet

• 

  RyanMorrisQ40

  —

  说人话、办实事，形容得贴切。

---

• 

  FrederickDavidson

  —

  豆包 yyds，日常离不开了。

---

• 

  毛浩

  —

  回不去了，已经成主力助手。

---

• 

  段雪_1

  —

  三强里豆包移动端体验最顺，响应快、中文自然，语音对话的节奏跟真人差不多。付费专业版目前坑有点多，桌面 Agent 还没打磨稳，但免费版日常完全够打，我打算先一直用免费版，等它把任务稳定性做好了再考虑升级，不着急交这个钱。

---

• 

  海浪382

  —

  退款这块真劝退，买专业版发现不好用想退，条款写「虚拟商品原则上不支持退款」，客服也一直拿这条挡。后来自己查了下，说是这种格式条款大概率会被认定无效，真要较真得走投诉或者起诉，可普通用户哪有这精力和时间跟它耗，只能认栽。

---

• 

  Margaret_CampbellII

  —

  专业版额度消耗快还不透明，说没就没了，没预兆。

---

• 

  KOste

  —

  专业版那个办公任务模式，我让它整理 C 盘、清理内存，操作完 C 盘可用空间不增反减，原本 3 个多 G 现在只剩 2 个多 G。理解力感觉不太行，还得我反复调教才能完成目标，这种活我现在是不敢随便交给它了，怕越弄越乱。

---

• 

  Jason.Wood3694

  —

  我冲着专业版 68 块一个月的桌面 Agent 去的，想让它根据 Excel 帮我出月报。结果实测频频掉线，额度用完当前任务直接中断，还没法交给别的 AI 接力，只能干等刷新，这一等一周就过去了，体验一言难尽。

---

• 

  暖阳589

  —

  AI 播客功能我很爱，把长文章转成音频，通勤路上听省时间。语音识别准，四川话粤语都能聊，方言交互这块豆包确实比同价位几个都强，我家老人用语音问天气、查菜谱也完全没问题，门槛真的低。

---

• 

  NMorales_7

  —

  生图偶尔抽风，但新版本全模态上来后稳定多了。

---

• 

  EHill_99

  —

  语音输入是真舒服，延迟基本感觉不到，做饭时候问它菜谱很顺手。

---

• 

  JohnJimenez520

  —

  让我惊喜的是它能直接解析抖音、小红书视频链接，总结脚本或者生成脑图，做自媒体的人应该会很喜欢。一键产出带货文案也是真方便，字节系生态打通得太深了，从选题到文案到分镜一条龙，省了我不少时间。

---

• 

  gAMEfIgAMERsIMMONS

  —

  免费的确实够用，日常问答写短文案完全没问题，没必要一上来就开专业版。

---

• 

  Kevin_Thompson_2022

  —

  写长文超过 800 字就开始跑偏，逻辑会散，长内容还是得用 DeepSeek。

---

• 

  BeverlyCampbellJr

  —

  用了三个月，辅导孩子功课确实省心，拍题就能一步步拆解解题思路，还会按我的要求出结构化练习题，数学的幂的乘方、语文的文言文断句都能讲明白，老母亲感动，至少不用自己硬着头皮去讲了。

---

• 

  GMoore_X1

  —

  深度推理和复杂代码还是不如 DeepSeek，但日常办公、内容创作、图像处理它更顺手。我的用法很明确：写代码跑数据开 DeepSeek，写周报做 PPT 刷短视频文案就开豆包，两个都免费，各取所长不冲突。

---

• 

  姚飞

  —

  做 PPT 确实快，给个主题几分钟出初稿，改改就能用，打工人福音。

---

• 

  redgoose402

  —

  AI 修图几乎无对手。

---

• 

  秦杰_1

  —

  跨平台操作偶尔触发别家 App 的风控，敏感场景得小心。

---

• 

  RaúlMartínez

  —

  界面干净没广告，基础功能永久免费，这点比很多套壳 AI 强。

---

• 

  DeFiGuru801

  —

  学生党表示：智能体搜索键被取消后找学习助手太麻烦了，希望官方恢复一下。

No reviews yet