Meta AI Chief Alexandr Wang: Team Creates Muse Model
Meta's Chief AI Officer Alexandr Wang stated that Muse was able to launch in its current form because the team meticulously reviewed and weighed every product detail and model choice before finalizing it. He described these individuals as capable of tackling hard problems and persisting through dead ends, expressing his fortune to work with them.
Muse is a personal AI agent launched by Meta, powered by the Meta Superintelligence Lab's model family, initially available on iOS, Android, and web in the U.S., with plans to integrate into smart glasses. The product operates through chat threads, allowing users to name, create avatars, and customize speaking styles; the agent works in a cloud-based virtual machine with a visible browser, capable of connecting to email, calendars, payments, and daily applications, with sensitive operations requiring prior confirmation.
There are three pricing tiers: a free tier covering most daily tasks, and paid tiers at $20 and $100 per month for heavy users to cover computing power. The product does not display ads, and the company claims to be exploring e-commerce revenue sharing. Checkout is processed through Stripe's Link, claiming to handle bookings, purchases, negotiations, form filling, and converting recipes into shopping lists. A sentinel system is also in place to review outbound actions from the virtual machine.
The same team previously delivered Muse Spark, Muse Image, Muse Video, real-time transcription, and the open-source 30 billion parameter Muse Glimmer. Starting with Spark 1.1, the model interface was opened to developers, with early adopters including Replit, Box, and Cline; version 1.3 emphasized programming and agent capabilities. Spark is positioned as a reconstruction following Llama 4, which was internally referred to as merely an "appetizer."
After setbacks with Llama 4, Mark Zuckerberg invested approximately $14 billion to $15 billion into Scale AI and brought in its founder to lead the lab. The lab initially operated with about 100 people in an independent workspace in Menlo Park. The company's capital expenditure guidance for 2026 is between $125 billion and $145 billion, aiming for over 1.3 million GPUs. The stock price rose about 6% on the day of Muse's release.
In market mechanics, the buyers are U.S. consumers who must provide email and payment permissions, while the seller is Meta, subsidizing free computing power through its advertising empire. This is event-driven: after social advertising growth peaked, it is necessary to shift conversations from Q&A to executable agents to lock users into new entry points. Funds flow from advertising profits to GPUs, high salaries for talent acquisition, and independent virtual machines; beneficiaries are platforms receiving subscriptions and future e-commerce commissions, while those under pressure include OpenAI, Anthropic, which charge per call, and users who must decide whether to hand over the keys to Meta.
Source: Public Information
ABAB AI Insight
Alexandr Wang started with Scale AI, turning data labeling into a supply chain for cutting-edge labs, and then at Meta, he invested around $14 billion to gain entry and control over the lab. Scale had previously supplied data to OpenAI and Anthropic, and after moving to the buyer side, he transferred the same delivery discipline to TBD Lab: nine months to redo infrastructure, architecture, and data pipelines, first proving capability with Spark, then stocking with Image, voice, and open-source Glimmer, with Muse being the model integrated into the consumer end requiring email keys. The approach is consistent with his strategy in the labeling business—first turning dirty work into a production line, then selling the production line as a platform.
Capital mobilization occurs on three levels: Zuckerberg uses cash to buy talent and time; the lab invests its budget into millions of GPUs and isolated virtual machines, rather than continuing to iterate on the Llama brand; on the product side, the free tier exchanges permissions, while the $20 and $100 tiers cover inference costs, with ads taking a back seat and e-commerce filling in later. The strategic motive is to transform superintelligence from a benchmark list into a daily agent that "spends your money and sends your messages," because only by touching payments and calendars can advertising companies turn social graphs into action graphs.
Similar comparisons include OpenAI's Operator and ChatGPT agents, Google embedding Gemini into search and Android, and Amazon transforming Alexa into a shopping assistant. Meta is in an expansion phase, shifting from model supplementation to entry point capture: Spark acknowledges that coding still lags behind Claude, yet it is ahead in deploying a 30 billion local model and sinking applications for volume. The industry position is no longer about publishing papers, but about who can get a billion people to plug their keys into their cloud machines first.
The structural judgment indicates a transfer of pricing power. Labs selling chat services by token have pricing power in the model; once an agent can place orders and change itineraries, pricing power shifts to platforms that own identities, relationships, and payment channels. The mechanism is: as computing costs are diluted by advertising profits, the free tier can fall below the marginal cost of pure API companies, with users trading privacy for convenience, and platforms exchanging convenience for the next round of transaction taxes.
ABAB News · Cognitive Laws
The model is only a semi-finished product; the entry point is the finished product.
The free tier is not charity; it is the price for exchanging keys.
Whoever can spend your money owns your next step.