Meta is forcing a reckoning across tech's most profitable business models by building AI agents that bypass the screens and ad placements that have powered Google and Amazon for decades. The company's new Muse agent reached 730,000 downloads in its first five days after launching on September 8, 2026, surpassing ChatGPT to become the top free app on the U.S. App Store, while its open-weight Llama model has been downloaded over 1 billion times since 2023.

What Makes Meta's Muse Different From Other AI Assistants?

Muse is not designed to answer questions like ChatGPT. Instead, it performs tasks on your behalf. The agent connects to everyday apps including email, calendars, Instagram, WhatsApp, and Facebook to handle reservations, form entries, product comparisons, and purchases without requiring you to visit websites or see product pages. It operates in a dedicated virtual environment called "Muse Secure VM" with its own browser, and all outgoing communication is monitored by an AI safety system called "Sentinel." Payments are processed through Stripe's "Link," which generates a disposable card number for each transaction so stores never see your actual payment information.

The pricing structure reflects Meta's strategic shift away from advertising. The service includes a free tier and paid plans at $20 and $100 per month. Unlike Meta's core social platforms, the company has explicitly stated that conversations with Muse are not passed to its advertising system. This separation of ads from the agent represents a fundamental business model change for a company historically dependent on selling targeted advertising.

Why Are Amazon and Google Treating AI Agents as Existential Threats?

The threat posed by Muse and similar agents is straightforward: if an AI can complete your shopping, booking, and purchasing tasks through conversation alone, you stop visiting websites. You stop seeing product listings. You stop viewing sponsored results. Amazon's advertising revenue exceeded $68 billion last year, a figure entirely predicated on the assumption that people will look at product pages and see sponsored placements. When an agent does the shopping instead, those ads never enter your field of vision.

Amazon responded decisively. On September 20, 2026, just 12 days after Muse's release, the company began blocking access to the agent, citing violations of its Terms of Service. This was not Amazon's first defensive action. The company has long shut out external agents from its platform, suing Perplexity's "Comet" browser and blocking shopping agents from Google and OpenAI. While Amazon won a preliminary injunction against Perplexity in March 2026, the Ninth Circuit Court of Appeals overturned that ruling on August 4, determining that users themselves, not AI companies, are responsible for accessing Amazon's computers under federal anti-hacking laws. Amazon's remaining weapons are contractual restrictions and terms of service enforcement.

Google faces the same threat but has chosen a different defensive strategy. Rather than block external agents, Google is dismantling its own cash cow with its own hands. At Google I/O in May 2026, the company announced that it would integrate AI agents directly into the search bar, allowing users to complete purchases, check ticket availability, and manage schedules without leaving search. Nick Fox, Senior Vice President of Search and Ads, positioned this as the first redesign of the search bar in 25 years. For paid Ultra members, Google has begun offering a "search agent" that monitors information in the background 24/7. By building agents into search itself, Google aims to keep users within its ecosystem and maintain control over the customer relationship, even as the nature of that interaction fundamentally changes.

How Do Meta's Open-Weight Models Create a Different Competitive Landscape?

While Muse represents Meta's direct challenge to Amazon and Google, the company's Llama model represents a different kind of disruption. Llama is an open-weight model, meaning Meta makes the underlying numerical weights that form the model's "brain" available for anyone to download and run on their own servers. This differs fundamentally from closed API models like Google's Gemini and OpenAI's GPT, where the weights remain locked within the provider's data centers and users only receive answers through an internet connection.

The distinction matters because it shifts control and reduces dependence. With an API model, the provider controls pricing, can change specifications, and can terminate service at any time. With open weights, users download the model once and can run it indefinitely on their own infrastructure, customize it with their own data, and avoid sending sensitive information outside their organization. Since Llama's debut in 2023, Meta has allowed free downloads and provided tools to modify the model. In March 2025, Mark Zuckerberg announced that cumulative downloads had surpassed 1 billion, representing a 50 percent increase from 650 million downloads just three months earlier. Companies including Spotify, AT&T, and DoorDash have adopted Llama for their own AI applications.

How to Understand the Three Models of AI Distribution

• API Provision (Closed): The model's weights remain stored within the provider's data center. Users send questions over the internet and receive only answers, paying for what they use. Google's Gemini and OpenAI's GPT operate this way, giving providers control over pricing, updates, and service continuity.
• Open Weights: The numerical weights are made public for download, allowing anyone to run the model on their own servers or computers and even retrain it with proprietary data. Users maintain their own data security and cannot be shut down by the provider, though the training data and methods are often not disclosed.
• Open Source in the Strict Sense: In addition to weights, the training data, procedures, and all documentation are made public with no usage restrictions. Almost no major companies pursue this approach due to competitive and safety concerns.

Meta's strategy of distributing Llama as open weights has created an ecosystem where companies can build AI applications without depending on Meta's infrastructure or paying per-query fees. This approach has made Llama the foundation for countless AI projects across industries, from healthcare to finance to e-commerce.

The broader implication is that Meta, Google, and Amazon are operating on a shared premise: the era of humans looking at screens to make decisions is ending. Google is rebuilding search to survive that transition. Amazon is building walls to defend against it. Meta is attacking from multiple angles, using both direct agents like Muse and foundational models like Llama to reshape how people interact with technology and make decisions. Which approach will ultimately prevail remains uncertain, but the competitive intensity suggests that the next phase of AI's impact on business will be defined not by model capability alone, but by who controls the interface between users and the digital world.