Ark Agent Plan

• 1. overview (634 chars)

  Ark Agent Plan is a subscription-based large model service package launched by Volcano Engine on May 11, 2026. It is officially defined as "the industry's first Agent package." It has expanded on the original Ark Coding Plan. It no longer only covers text models for programming, but multi-modal models such as text, code, images, videos, and vectorization, as well as Harness tool chains such as network search, memory, and Auto scheduling, are packaged into the same subscription. It also introduces a unified points billing system called AFP (Agent Fuel Points, Agent Fuel Value). It is aimed at domestic developers who use tools such as Claude Code, OpenCode, TRAE, OpenClaw, Hermes Agent, etc., and want to use domestic models for Agent development.

• 2. Product background (992 chars)

  Ark Agent Plan comes from Beijing Volcano Engine Technology Co., Ltd., which is the Volcano Engine owned by ByteDance, and is part of its "Volcano Ark" large model ecosystem. Volcano Ark itself is a platform-level service and the base of the entire large model ecosystem, carrying the capabilities of model access, fine-tuning, deployment, and multi-model collaboration. Agent Plan is a layer of subscription products that grows on this base, alongside "Ark Coding Plan", "Saving Plan" and "AgentKit".   The evolution logic of this product line is very clear. Volcano Engine first used Coding Plan to open up the situation of "subscription manufacturing and selling programming tokens", relying on Byte's own resource guarantee and engineering capabilities to keep the price very low (limited-time promotions starting at 9.9 yuan have always been its signature). Agent Plan is an upward expansion on this basis. In official words, it is a key step from "selling models" to "selling Agent productivity". The release time is chosen around the 2026 Force Prime Minister Conference (spring), and it can also be seen that Volcano Engine wants to promote it as a key product of the year.

• 3. Product functions and usage experience (2473 chars)

  The simplest formula to understand Agent Plan is: Agent Plan = Coding Plan + image generation + video generation + other tool chains. Coding Plan can only cover the tokens consumed in the coding process, while Agent Plan, in addition to programming tokens, can also call drawing (Seedream 5.0 lite), video generation (Seedance 2.0), Internet search, embedding and other capabilities, which is equivalent to collecting AIGC services that originally required each company to receive APIs separately into one subscription.   The model lineup is one of its big selling points. Byte's self-research includes Doubao-Seed series for programming, Doubao-Seedance 2.0 for video, Doubao-Seedream 5.0 lite for images, Doubao-embedding-vision for multi-modal memory and vector retrieval; third-party mainstream models are connected GLM-5.1/5.2 (Wisdom Spectrum), Kimi-K2.6 (Moonshot), MiniMax-M2.7/M3, DeepSeek-V4 series, etc. Users can freely switch between these models, or directly turn on the Auto mode, allowing the system to automatically match the most cost-effective model according to the task, eliminating the trouble of manual selection.   Even more noteworthy than the model is the Harness tool layer. For the first time, Agent Plan deeply integrates the "model" and the "Agent tool chain" for out-of-the-box use: Internet search is a real-time retrieval capability that has the same source as Doubao, and there is no need to build a search interface by yourself; the memory capability is driven by the built-in embedding model, supporting long-term memory; the Auto mode is responsible for task scheduling; multi-modality is to uniformly schedule the generation and understanding of text, code, images, and videos in the same set of APIs. In terms of ecological compatibility, it is adapted to programming tools such as Claude Code, Codex CLI, TRAE, Cursor, OpenCode, Cline, Roo Code, and Kilo Code, as well as autonomous agents such as OpenClaw and Hermes Agent. The same package can be shared among multiple tools, but the quota is shared, and the consumption of all tools is deducted from the same package.Subscribing to Medium and above will also give you an ArkClaw lightweight version instance for free. ArkClaw is a cloud-hosted version of the agent built by Volcano Engine based on OpenClaw. It features zero threshold, no operation and maintenance, and is online 24/7. It can be connected to messaging channels such as Feishu, DingTalk, and corporate WeChat, and supports scheduled tasks, personality configuration, and skill plug-ins. The difference between it and the open source self-hosted OpenClaw lies in the word "hosting": OpenClaw requires its own environment, API Key, server and operation and maintenance, while ArkClaw can be deployed in the cloud with one click, automatically upgraded, faults can be repaired with one click, and tasks can run in the background even when the computer is turned off.

• 4. Pricing and business model (1596 chars)

  The Ark Agent Plan for individuals is divided into four levels, with monthly payment: Small 40 yuan, Medium 200 yuan, Large 500 yuan, and Max 1,000 yuan. The billing unit is AFP, and the official caliber is that 1 AFP is approximately equal to 10,000 tokens. The corresponding monthly quotas for the four tiers are approximately Small 20,000 AFP, Medium 100,000 AFP, Large 250,000 AFP, and Max 500,000 AFP. The usage is released on a rolling basis of 5 hours, weekly, and daily. Inventory is released at 00:00 every day and is subject to availability, so the popular tiers have a bit of a "rush purchase" attribute.   The AFP deduction coefficient depends on the model: the ultra-fast model is about 0.5 AFP/10,000 tokens, the standard model is about 1 AFP, the advanced model (code/pro, etc.) is about 5 AFP, and the third-party model (GLM, Kimi, etc.) is about 9 AFP; image generation is about 99 AFP/picture, and video generation is between 36 and 230 AFP depending on the model and input type. The advantage of this design is that developers only need to keep an eye on AFP consumption and do not need to convert back and forth between multiple billing systems. They know the budget at the beginning of the month and will not easily overspend at the end of the month.   In terms of gear selection, the consensus among all evaluations is relatively close: Small 40 yuan is suitable for individuals who are light adopters and pure text programming, but does not need video generation; Medium 200 yuan is the main gear for daily Agent development, and is also the sweet spot for unlocking video creation and smart partners; Large and Max are for heavy multi-modality, long processes, and team scenarios. In addition, there are four levels from Team Small to Team Max for enterprises, which support purchase by seat and sub-account allocation, starting from 5 seats, and promise not to record user requests and model return data.

• 5. User reviews (1166 chars)

  Positive feedback focuses on the matter of "aggregation". Many developers believe that the biggest value of the 40 yuan basic version is not that it is cheap, but that it opens up the calling channels of multiple models such as GLM-5.1, Kimi-K2.6, etc. at once, eliminating the trouble of registering multiple platforms separately, and it is completely sufficient for text-based Agent development. Users who make videos value the 200 yuan price tag more. After unlocking Seedance 2.0 and working with OpenClaw, they can build a nearly fully automatic video production line. The three-piece set of direct domestic connection, no need for an agent, Auto mode plus Internet search and memory is ready to use right out of the box. These features have a good reputation among terminal parties. A third-party reviewer gave it a score of 9.6 for "domestic experience".   The negative feedback is also very real, mainly about the price structure. Compared with the old version of Coding Plan, the token unit price of Agent Plan has been increased by 2 to 2.5 times; what is even more worrying is that long contexts exceeding 128k will trigger a multiplier of up to 7.5 times, which is very unfriendly to users who like to feed the entire repo at once. Some people also complained about the high learning curve of the terminal UI, and the fact that it is sold out at 00:00 every day and requires time to place orders.

• 6. Industry evaluation (1184 chars)

  The media generally regards Agent Plan as an upgrade of the Volcano Engine business model. Sohu, Renrenkaika and other media position it as "the industry's first AI capability package", emphasizing that it has raised the threshold of Agent development to a level within the reach of mass developers. CodePick’s in-depth interpretation gives a more calm judgment: Coding Plan is still the best cost-effective solution for pure programming scenarios, while Agent Plan is a new choice when “multi-modal plus Agent tool chains are required”. The two are complementary rather than substitutes. Test bloggers such as Sumi (xmsumi) believe that the Medium quota of 100,000 points/month is basically sufficient for daily use. The consumption of tens of AFPs for a picture and hundreds of AFPs for a video is within an acceptable range.   Looking at the competitive landscape, half of Agent Plan’s targets are overseas and half are domestic. Externally, it is a domestic transit alternative to Claude Code subscriptions. Its selling point is domestic models plus domestic direct connection; internally, it forms a high-low match with its own Coding Plan, and also competes head-on with the subscription-based services of other domestic model platforms. The ArkClaw line directly competes with Feishu AI and open source OpenClaw, and has an advantage in the dimension of "hosted operation and maintenance-free".

• 7. Controversies and Risks (945 chars)

  The biggest controversy is the price increase. Raising the token unit price by 2 to 2.5 times and adding a 7.5 times multiplier to long context means that it will be more expensive for some heavy users who originally used the Coding Plan to migrate there. This makes the propaganda of "increasing volume without increasing price" untenable under certain usages. Although the mechanism of releasing inventory every day and while supplies last can prevent resources from being occupied by a few people, it is an uncertain factor for teams with fixed budgets and who want to stabilize usage.   Another hidden concern is the binding ecology. Agent Plan packages models, tool chains, and smart partners together. It is indeed convenient, but it also means that the more users use it, the more they rely on the Volcano engine, and migration costs will increase over time. For data-sensitive enterprises, although the official commitment is not to record request and return data, but the entire link of programming, creation, and office work is handed over to the same cloud platform, they still need to do their own compliance assessment.

• 8. Applicable groups and usage suggestions (806 chars)

  If you are a domestic developer who wants to play Agent and multi-modality at the same time in the terminal, especially a user of tools such as Claude Code and OpenCode, but do not want to bother with agents and multi-platform registration, Agent Plan is worth a try. For teams that make content-based products and need to add images and videos while writing code, the Medium file is a more suitable starting point. Enterprise-level, multi-seat heavy teams can look directly to the Team series.   On the other hand, if you only do pure text programming and have no need for multi-modality, then the old Coding Plan is still more cost-effective, and there is no need to spend more for capabilities that you cannot use. Users who like to feed a large repo to the model in full at one time should pay special attention to the 7.5x magnification of long context. It is best to run a real task in Small mode for a week to figure out the AFP consumption before deciding whether to upgrade.

• 9. Summarize (546 chars)

  Ark Agent Plan is a key move for Volcano Engine to shift from "selling models" to "selling Agent productivity". It connects multi-model, multi-modal and Agent tool chains with one subscription. It is an integration solution with a lower threshold for domestic Agent developers. However, the price increase and long context magnification also make it more suitable for multi-modal scenarios rather than pure programming heavy users. Overall, it has established the "Agent package" category first. How far it can go in the future depends on whether the quota, price and ecological compatibility can continue to keep up with developers' real usage.

No in-depth report yet

• 

  TerryScott562

  —

  40块一个月是真便宜。

---

• 

  Catherine.Perry_7

  —

  最大的爽点其实不是便宜，是一个订阅把 GLM、Kimi、DeepSeek 全打通了，再也不用在几个平台之间来回注册切 key，光这一点就值回票价。文本类 Agent 我做下来完全够用。

---

• 

  Susan_RussellII

  —

  AFP 这个积分制说实话比按 token 计费直观多了，月初心里有数，月末不至于爆预算，控成本友好~

---

• 

  Abigail_Foster_99

  —

  劝一句纯写代码的别急着冲。我从 Coding Plan 迁过来发现单价涨了两倍多，最坑的是超 128k 长上下文直接 7.5 倍倍率，我习惯把整个 repo 怼进去的，一算账反而更贵了。多模态需求不强的话老老实实待在 Coding Plan。

---

• 

  orangekoala695

  —

  在 Claude Code 里配了一下，国内直连不用挂代理这点太舒服了。

---

• 

  cxxt4bruw

  —

  视频生成那额度是真能烧，跑几条 Seedance 燃料值哗哗掉，希望官方优化下消耗。

---

• 

  AdrianaSánchez

  —

  用 OpenClaw 接上之后我就丢了一句「给科技博客写篇 AI 趋势的文章，带标题正文配图」，它自己调 GLM 写文案、调 Seedream 出图，一条龙下来了，这种全自动的感觉确实上头。

---

• 

  LindaFarragher

  —

  Small 档不能生视频这点没看清，买完才发现，想玩多模态的直接上 Medium 吧。

---

• 

  tiMIL

  —

  每天 00:00 放库存售完即止是什么抢购机制啊，热门档还得掐点下单，救命。

---

• 

  CarlaAcosta

  —

  Medium 及以上送的那个 ArkClaw 智能伙伴我一开始没当回事，用了几天真香，7×24 挂着帮我盯任务、做纪要，不用自己搭 OpenClaw 折腾环境，白嫖的东西反而成了我留在这档的理由。

---

• 

  meWAG

  —

  问一下，同一个套餐能在 TRAE 和 OpenCode 里一起用吗？

---

• 

  Frank127

  —

  本质上就是从卖模型 token 变成卖 Agent 燃料，把搜索、记忆、图像视频这些原来要自己拼的能力打成一个包，省了不少胶水代码。思路是对的，就看后续额度和价格能不能扛住真实用量了。

---

• 

  安然561

  —

  终端 UI 对新手是真不太友好，配置那一段折腾了半天，文档能再详细点就好了。

---

• 

  qa_5s0kc

  —

  对比 ChatGPT Plus 一个月 20 刀，这个 40 块的全家桶确实是目前国内最便宜的一档了，自媒体和产品经理拿来搜资料写文案做竞品分析绰绰有余。

---

• 

  BrodyTorres

  —

  多模态 Agent 玩家狂喜，编程加生图加生视频一个订阅全覆盖，泰裤辣。

No reviews yet