In-depth discussion of Personal Agent: Muse costs $10 to $100 per month, and the industry awaits the iPhone moment

[[**Personal Agent is shifting from chatbots to proactive cloud-based assistants that get things done, with shopping and personal finance the first scenarios to prove value; but per-user monthly costs reach $10 to $100, and traffic gateways and business models may be restructured, with the industry still waiting for its "iPhone moment."**

**Key points**

**1. Proactive task execution and cloud hosting lower the barrier**

After accounts are connected, the agent can proactively handle affairs; Muse and others productize cloud environments, connectors, skills, and multi-agent collaboration, so ordinary users need not configure and maintain things themselves.

**2. Shopping and finance prove value first, but retention remains a hurdle**

Restocking, price comparison, tax filing, and insurance save time with intuitive benefits; Muse has about 3 million DAU, but its day-14 open rate is lower than ChatGPT's, and about 90% of interactions are still initiated by users.

**3. High costs put the internet charging model under pressure**

Muse and Instinct cost about $1 to $5 per user per day, and early teams estimated $50 to $100 per month; free services need advertising, transaction revenue sharing, or cost reduction to cover costs, and subscription growth is difficult.

**4. Big tech holds advantages in compute, identity, and gateways**

Meta and Google have advantages in accounts, models, compute, and device gateways, while Apple can serve as a neutral platform; independent teams rely more on product momentum, and domestically they are constrained by the closed ecosystem of super apps.

**5. Early stage is product-focused, mid-term is model-focused, long-term is ecosystem-focused**

This article is from the WeChat public account: Overseas Unicorn, author: Best Ideas Community, cover image from: AI-generated

Personal Agent is the liveliest area of AI in Q4 this year: Meta's Muse, OpenAI's dots, Instinct, and Anthropic's newly launched Claude Duet have put big tech, AI Labs, and startups on the same stage.

It is also the AI proposition most like the internet: if users start handing their needs to agents, then everything from user gateways and traffic allocation to transaction loops will be reorganized.

However, given that the per-user service cost of Personal Agent is far higher than that of traditional internet products, it may also fail from a product or business model perspective.

Today's imagination about Personal Agent will inevitably be limited by our past cognition and experience. Looking at the evolution of the PC, from Palm and Windows Mobile to the iPhone, only then was a mobile computing experience truly suited to the masses found. Personal Agent has only just begun, and the whole field likewise needs an "iPhone moment," that is, finding a personal intelligence product form that truly lets ordinary people use it naturally and continuously.

Below is a summary of our views from the Best Ideas discussion on the Personal Agent theme.

1. What is new about this round of Personal Agent?

1. Compared with chatbots, "more proactive and able to get things done" is the key advance of personal agents. Past interactions were mainly users asking questions and AI providing information; but after connecting to services such as Gmail and obtaining user context and authorization, agents began to have the opportunity to proactively handle affairs for users.

2. The relationship between personal agents and chatbots may be more like the relationship between recommendation and search. The appeal of Personal Agent to users lies in proactivity. In chatbots, many users do not know what they can ask AI to do, and it is hard to express their needs clearly. If users must ask first every time, usage will be limited, but when the Agent has proactive capabilities, this problem can be effectively solved.

3. The explosive popularity of OpenClaw earlier this year and personal agents are part of the same narrative; the most important advance of personal agents is productization and cloud hosting, which solves the burden on ordinary users of building, configuring, and maintaining things, making them more "ready to use out of the box."

4. For example, Meta's Muse actually puts a continuously running cloud environment, connectors that connect accounts and services, reusable skills, and background multi-agent collaboration into the same product. Users can directly assign tasks to Muse on their phones without having to configure these components one by one first.

5. Meta AI product head Nat Friedman also explicitly mentioned that Muse's product design drew considerable inspiration from OpenClaw.

6. The product changes that model progress brings to personal agents are very important: advances in models' capabilities in long-task execution, tool calling, and instruction following have made personal agents begin to become usable. If future models with tens of billions or two hundred billion parameters can handle long tasks well, they may bring personal agents user growth like what GPT-4o brought back then.

7. Grok Bot and Muse cannot be compared head-to-head:

Grok Bot is more like a work assistant for one-person companies (OPC) or small US businesses. It can operate applications, browse the web, and call tools to handle specific tasks for business owners or employees. Users can also connect tools such as Codex to the cloud computer hosted by Grok Bot, after which they only need to assign tasks uniformly to Grok Bot, which calls the corresponding tools to execute them.

But personal agents, especially Muse, clearly, does a better job optimizing for real-life scenarios. In some users' firsthand experience, Muse can complete tasks such as shopping, haggling with sellers on Facebook, and insurance within 10 minutes, while other agents often take 20 to 30 minutes.

8. As the functional boundaries of these two products keep shifting, what the two can do has also come to overlap considerably. For example, Muse also launched an enterprise edition last month, adding connectors and skills for small businesses to the Muse App, positioned as an AI business assistant that can read operational data and execute tasks across software.

9. If a personal agent can do well on voice, cross-device use, proactive interaction, and optimization for personal scenarios, it can still make the product feel very different to use.

II. How the products feel to use

10. From users' direct experience, people generally use personal agents for repetitive, cumbersome information processing and retrieval tasks, among which shopping and personal finance tasks are the scenarios where the value of personal agent products is most keenly felt.

11. "Regular restocking" may be a relatively clear scenario for personal agents today: the fresh groceries bought each week are similar, the goal is clear, and placing repeat orders is time-consuming, so it can be handed entirely to an agent.

12. In addition, the information retrieval and comparison that AI excels at can also be externalized into the shopping experience. For example, one user was used to buying items on Amazon, but the agent found Best Buy, which the user had not previously considered, and ultimately led to a purchase.

13. Personal finance-related: for example, tax filing, insurance, and product subscription management. The typical feature of such tasks is that the process is cumbersome, but because real money is at stake, users previously had to spend a lot of time handling them. Moreover, in such scenarios, because a personal agent can both save users time and deliver directly perceptible benefits, this scenario has greater room for imagination than ordinary shopping scenarios.

14. Based on the above, from a developer's perspective, to judge whether a task is worth handing to an agent, one should not only look at the cost of the agent executing the task, but also compare how much time and hassle it saves the user in communication and paperwork.

15. The product experience mentioned above is mainly for overseas markets. Domestic personal agent products first need to face the access problem: closed ecosystems and an insufficiently mature API ecosystem make it harder for independent agents to obtain account data, and achieving subsequent operations on behalf of users is an even greater challenge.

16. As a typical To C application, the product persona of a personal agent may also be one of the factors affecting whether ordinary people are willing to try a personal agent. Muse feels more approachable, making people who are unfamiliar with AI more willing to try it. Instinct, based on iMessage, has no clear persona, while the personas of Grok Bot and ChatGPT's dots lean toward pixel or geometric design.

III. Personal Agents need their own "iPhone moment"

17. AI products have already accumulated a huge user base. After deduplication, products such as ChatGPT, Claude, and Gemini are estimated to have perhaps 600 million daily active users, already a fairly meaningful number, while today's leading products such as Facebook are roughly at 2 billion DAU.

18. This trend is most directly reflected in the change in traffic allocation: when users hand their needs to an agent, which searches across platforms, compares, and completes purchases, the range of choices previously limited by personal attention and usage habits may expand. For example, a user may originally only be used to shopping on Amazon, but the agent may find channels such as Best Buy that were not previously considered. This change may also squeeze some businesses that profit from information asymmetry.

19. Judging from Muse's early user scale, a "Facebook-level opportunity" is also foreseeable: by estimation, more than a month after launch, Muse already had ~3 million DAU, with each daily active user initiating more than 10 prompts per day. Although traffic referrals from Meta's apps contributed ~60%–70%, Muse's data is still an important signal validating demand for Personal Agents.

20. But the real-world bottlenecks of Personal Agents are also obvious:

Today users may still need to use Muse for 2-3 weeks before gradually entering a state of heavy usage;

One possibility is that ordinary people simply do not have that many things worth handing to a personal agent, that is, how large the population is that can provide background such as email, professional materials, and code repositories over the long term while continuously proposing high-value tasks still needs discussion.

21. But Personal Agent-type products have also begun to address the matter of helping users "discover needs." Take Muse as an example:

Muse uses Goals to save users' long-term goals, and then uses ideas to provide new task suggestions, with the aim of hoping that after finishing one task, users can still find the next thing worth assigning.

Some research estimates that currently only about 10% of interactions are initiated by the agent first (when the agent initiates proactively, it is mostly based on the user's recent plans, visited locations, or search history), while 90% are still initiated by users first.

22. Of course, the data users display may not reflect real demand. From recommended content to proactively completing actual actions, the difficulty facing products is higher, so the realization of proactivity will be a gradual process.

23. However, there is still a voice questioning whether some of the current market enthusiasm for Personal Agent comes from the public's imagination of personal assistants, and by contrast, agent colleagues in enterprises may be more likely than consumer-grade personal agents to produce trillion-dollar companies:

Enterprises have clearer sources of tasks;

An enterprise's business records, internal communication information, and new items in processes can all provide the agent with things to handle;

As long as a few people in the team can propose valuable tasks and build processes, they can drive other members to use it, rather than requiring everyone to constantly come up with new tasks.

24. People often imagine the future along existing product forms: today's imagination of Personal AI may be like trying back then to move the PC onto a small screen, equipped with a keyboard and stylus, yet not directly achieving a mobile computing experience truly suited to the masses, until the iPhone appeared and realized this goal. This evolution also shows that the process of finding a product form that lets ordinary people use Personal AI naturally and continuously may be longer than expected.

25. The impact of a long time window is not that laggards will be treated more leniently, but that it makes leaders extremely tormented as well: they need to build ecosystems, continuously retain users, and even persist in investing while making losses, ultimately crossing the inflection point.

IV. What does improving product capabilities rely on?

Models

26. What is worth watching next is whether models can continue to lower the threshold for using personal agents, making them noticeably more usable like coding agents.

27. Designing benchmarks around personal tasks and then tuning models based on performance can greatly improve the agent's performance on these tasks. AutomationBench in office scenarios provides a reference: in a simulated enterprise environment, let the agent complete cross-application tasks through APIs, then check whether the required operations actually take effect.

28. Meta's agent model Muse Wonder strongly emphasizes active memory and active focus capabilities, corresponding respectively to retaining important context across conversations and continuously executing around the goal in long tasks.

29. Muse currently still uses conventional tools such as search to read memory, and provides relevant content as needed before model inference. If the work of retrieval and information provision has already been done well, what improvement can specially training the model's memory capability still bring?

Context

30. Ordinary users may not care which model is used behind the scenes, and care more about how the product works. People are beginning to discuss product experience more, and the importance of the product itself is rising.

31. Some users also believe that in the long run, obtaining users' personal information is more important than using SOTA models, and that models may even be the least important link in a personal agent.

Using a computer as an analogy, changing the CPU will change the running speed and the software that can be used, but as long as the original data and state remain, users will still feel this is their own computer.

Therefore, compared with coding tools that leave the main outputs and project state in the code repository, a personal agent, in addition to leaving the results of a single task, also needs to preserve user data and task progress over the long term, and have accounts and identities that can use external services.

32. But the agent being able to obtain the user's background information does not mean it understands what result the user wants. For example, some users found when switching between Claude Code and Codex that even if the background information the original agent had mastered was moved over, the new agent might still not understand the user's real needs.

33. Even if the agent understands what result the user wants, when handling matters the agent still needs to distinguish which actions it can complete on its own and which must first be asked of the user.

For example, reading emails, drafting replies, and directly sending emails bring different consequences.

Writing code can be checked by running and testing, but personal affairs often do not have the same clear way to check. If every step requires user confirmation, the user still has to spend effort watching; if the agent is allowed to handle everything on its own, the agent may send the wrong message or make promises that are hard to retract.

34. To enable the agent to do professional tasks well, there is also a division-of-labor idea: a main agent that holds personal identity, memory, goals, and permissions calls professional agents such as programming, legal, health, or financial ones as needed.

For example, in programming tasks, some developers use DeepSeek V4.1 Flash to drive Hermes, and install Codex CLI on the same computer; when code needs to be written or bugs fixed, they have Hermes first write a task ticket, then hand it to Codex for execution.

35. Personal agents focus on the product in the early stage, on the model in the middle stage, and on the ecosystem in the long term.

Ecosystem

36. The ecological environments at home and abroad differ greatly, and this difference limits the gap between domestic and overseas Personal Agent product experiences:

Domestic personal information, payments, and life services are more concentrated in super apps, and independent products have greater difficulty reading the personal information within them.

Many services in the United States are still built around PC web pages, and agents operating web pages through browsers can already provide a good experience.

37. But even if the agent can technically operate web pages, platforms may also restrict its access; for example, in September this year, Amazon blocked Muse, stop it from shopping on behalf of users.

One explanation is that platforms are unwilling to let new entry points intercept user relationships and business interests. Therefore, if agent service providers and platforms can negotiate a profit-sharing arrangement after cooperation, resistance to ecosystem access may decrease.

38. In addition to accessing large platform ecosystems, agents can also connect directly with merchants. For example, Luckin has already provided an MCP service that supports agents in searching stores, selecting products, and placing orders, while key steps still require user confirmation and payment.

39. The industry is also trying to establish a universal access protocol. Meta and Sierra are working with Instinct, Shopify, Stripe, Walmart, and other companies to develop the Personal Agent Protocol, letting users decide which permissions to grant agents and enterprises decide which operations agents are allowed to perform.

40. However, even open payment or agent interoperability protocols may not solve all access problems. When actually executing, agents still have to handle personal data access, account authorization, and how user devices and the cloud coordinate to execute tasks and protect privacy.

41. These ecosystem access challenges also give rise to a startup idea: can specialized service providers help agents access different platforms and obtain platform data access authorization? Perhaps big companies and startup teams can both try to provide such services, but whether it can become a business is still uncertain.

V. Cost Estimates

42. The cost of a personal agent includes model inference, the runtime environment, and the configuration of an external identity. Estimates of the service cost of a personal agent vary widely from person to person.

Some practitioners estimate that Muse and Instinct require about $1 to $5 per user per day.

Referring to Morgan Stanley's estimate, if a user uses 5% to 10% of their token quota each week, Muse costs about $10 to $18 per month to serve one user, of which the CPU sandbox cost is about $5 to $8.

43. After OpenClaw went viral, a team that had once tried to productize a personal agent shared the cost of building and running the product at the time:

Equipping each agent with an independent computing environment and a reliable network identity requires at least about $50 per user per month (excluding tokens): the team initially tried Mac Mini, found it expensive and difficult to scale, but after switching to virtual machines with a desktop, discovered that even at AWS large-customer prices, it still cost $10 to $15 per month. Under the team's plan at the time, the Gmail account used to register for external services also cost $14 per month, and adding identity configuration such as phone numbers and credit cards, fixed expenses were already not low.

Token expenses vary with the model type and user usage intensity. In the early days, the team used the best model available at the time for the personal agent, and calculations found that heavy users could consume about $1,000 per month in tokens alone.

Overall, when users normally use a personal agent product, including tokens and computing resources, it costs about $50 to $100 per user per month, which was already the lowest magnitude at the time.

44. In addition to cost, the team also encountered scaling access problems: some websites restrict registration, and the email, phone, and credit card identities required by large numbers of agents are also difficult to configure in a standardized way. Therefore, increasing the budget does not necessarily mean serving more users.

45. But when estimating costs, pay attention to distinguishing usage scenarios and resource allocation:

In enterprise scenarios, a 20-person finance team can share one virtual machine, and an 80-person product and research team can share two or one high-spec virtual machine.

Virtual machine resources can also be oversold, that is, the total resources allocated to multiple virtual machines can exceed the resources actually owned by the physical machine, and the proportion of CPU that can be oversold is higher than that of memory;

In addition, a work agent may run for more than ten to twenty hours a day, while an ordinary life assistant may have only a dozen or so requests per day, so the cost of a life assistant should not be estimated in the same way as a work agent running for long periods.

46. Considering that one reason personal agent products have concentrated recently is that more teams can already build models close to SOTA level. Therefore, vendors with their own models and computing power can also try to further control the operating cost of agents by distilling small models.

47. Ecosystem openness may reduce the execution cost of a single task, but more reliable fulfillment may also bring more tasks. After model prices fall, agents can also invest more inference in discovering needs, so a decline in unit cost does not mean a decline in total token consumption.

48. In addition to the cloud solutions commonly used today, some have also suggested that as local model capabilities and operating efficiency improve, some agent tasks could be handed over to local models on user devices. But this solution raises a question: exactly which tasks can be handed over to local models, and which still need to call cloud models?

49. Being able to run models locally still entails device maintenance costs for ordinary users, so services aimed at the general public may still be mainly cloud-based, while data-sensitive people or those with special needs will choose local solutions, and the two may coexist for a long time.

50. By 2027-2028, changes in technology, the supply chain, and market supply and demand may further lower the threshold for local deployment, making local agent devices easier to promote. By then, such devices may become standard products that can be purchased and used directly, like enterprise laser printers.

VI. What kind of charging model is more suitable for Personal Agent?

51. The biggest difference between AI applications and the mobile internet lies in marginal cost. The per-user service cost of a personal agent may be 2-3 orders of magnitude higher than that of the traditional internet, so the internet business model may fail.

52. Subscriptions are the charging method for most Agent products, especially productivity products, but Muse already offers a free service, so for personal agents, under today's competitive environment, it is very difficult to achieve growth through consumer subscriptions in the short term.

53. However, although subscriptions face competition from free products, they may be more in line with the starting point of a personal assistant saving users' attention.

54. But users who are willing to spend money to save time and value results may accept usage-based charging, or pay-for-results with a share of the proceeds:

For users: if an agent saves a user $1,000 in insurance premiums, then a fee can be charged from the savings;

For counterparties: the business model of an independent personal agent like Instinct can be summarized as "AI Ctrip," helping users connect with external service providers and complete service purchases, with the agent service provider then taking a share of the transaction.

But some question earning revenue through transaction revenue-sharing, because high-value intents with clear commercial purposes such as shopping account for only ~5% of ChatGPT queries.

Note: David Dugan, head of OpenAI's global advertising solutions, publicly stated in June 2026 that about 20% of ChatGPT queries have "direct commercial intent."

55. Muse offers a free life assistant service, which is very attractive to ordinary people; some users even compare Muse's emergence to the launch of Gmail back then: a personal assistant service that ordinary people could originally hardly afford long-term has the chance to become part of daily life.

56. However, whether Muse can offer free services long-term still depends on what revenue sources it can find. Some are optimistic about earning advertising revenue through Muse Feed, because when users use a life assistant, they may be more willing to accept ads and commercial recommendations. By contrast, when users use ChatGPT and Claude for work, they strongly dislike commercial recommendations inserted into answers.

57. But as mentioned earlier, if serving one user with Muse costs at least $10 to $20 per month, then even if Muse reaches the per-user revenue level of Meta's other mature apps, it may still not cover costs.

Meta's Q2 2026 financial report shows that the ARPP (average revenue per person per quarter) of Meta's apps is $16.86, equivalent to about $5.62 per month, including advertising and other revenue from apps such as Facebook and Instagram.

Therefore, Muse should perhaps first lower the monthly service cost per user by another order of magnitude, to about $1 to $2, so that revenue covers costs first, and then carry out model and product improvements.

58. Some also believe that the tasks users give to Muse can more directly reflect user needs, and if this information can help Meta increase revenue from its existing advertising business, the new revenue can be used to subsidize Muse's service costs.

Note: In the blog announcing Muse, Meta publicly promised that Muse will not share data from user conversations and virtual machines with the advertising system.

59. Meta and Google are favored first because of their advantages in identity authentication, foundation models, and user base, and more importantly because of their computing power reserves: in the Personal Agent competition, only vendors with sufficient computing resources can continuously provide free services and acquire users.

60. Company brand affects whether users are willing to connect accounts; for example, one Muse user said that compared with Instinct, he is more willing to hand account permissions to Meta.

61. Google's advantage comes first from Gmail: many formal matters overseas are handled through email, and the mailbox both preserves historical correspondence and can continuously bring new tasks to the personal agent.

62. Pixel and Android also give Google a device-layer entry point, enabling integrated hardware-to-software optimization of the user experience rather than relying only on software cooperation with phone manufacturers.

63. However, how many resources Google will invest in personal agents still depends on business priorities and computing power allocation:

Google's external customers, search recommendations, DeepMind, and cloud business are all using computing power, and consumer agents have lower priority;

But search is itself a step in an agent completing tasks, so personal agents are actually closely related to Google's main search business; if computing power supply increases in the future, Google may raise the priority of personal agents;

64. Smartphones represented by the iPhone are truly mass-market personal computing devices, continuously carrying users' schedules, locations, communications, and service records. Therefore, the role of device platforms such as Apple in the Personal Agent competition cannot be ignored.

65. A neutral platform role is also Apple's advantage:

As a platform, it may be easier for it to coordinate developers and service providers, enabling agents to access more services, and service providers may also be more willing to cooperate;

By contrast, Google runs its own apps and services and competes with some partners, so partners may be reluctant to hand it the user entry point; and the Android ecosystem is more open, making it harder for Google to uniformly require developers to cooperate.

66. But whether OS permissions have an irreplaceable role is still debated. Some also argue that Apple and existing phone and device manufacturers will not become the most important players in personal agents, because there are currently no tasks that can only be completed with OS permissions and that cloud agents cannot do. Moreover, apps are merely the current entry point through which users use agents. If in the future agents can proactively handle things for users, and users no longer need to use them through a fixed app, then the importance of existing app and device entry points may decline.

In response to this view, some added that a small number of identity verification scenarios still rely on physical devices: certain verifications need to be completed through physical devices at a specific time and place, and some website logins require scanning the face with the front camera or scanning a code with the rear camera, which are tasks that cloud agents cannot complete.

OpenAI

67. OpenAI, like the major internet companies, has a huge user base and frontier models, although the interaction experience of dots is not the best right now.

Independent teams

68. Those more optimistic about independent teams believe that startup teams are driven to build products by their own desire to do this well, while big companies may only follow up after seeing a business opportunity. In this camp's view, the product form of personal agents will continue to change, and teams need to keep investing and making product choices, so why they initially wanted to do this matters a lot.

69. There are also different ideas about where small teams should start.

Idea one: Professional needs can already be met by adding AI features to vertical tools, so there is no need to build another vertical personal agent.

Idea two: Personal agents can be combined with hardware, with hardware serving as the medium for understanding users' situations and for letting users access and use the product, which may provide unique user experiences in fields such as education and health. AI Venture Capital Daily Channel: Frontier Technology. This content is published with the author's authorization; the views represent only the author and do not represent Huxiu's position. If you have objections or complaints about this manuscript, please contact tougao@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]