Meta’s Muse did not reach the top of the App Store with the world’s best model. It did so with a product that could act, a network that could distribute it, and a position close to the transaction.
Two headlines in the same week pointed to the next phase of the AI market.
On 22 September, OpenAI and Anthropic released new models with a clear message: advanced intelligence is getting cheaper. OpenAI cut API prices for GPT-6 Sol and Luna by 50 per cent compared with the promotional prices of their GPT-5.6 predecessors. Sol targets demanding work such as coding and agentic workflows; Luna is designed for high-volume tasks where cost matters. Anthropic launched Claude Opus 5.5, saying it performs at the level of Claude Fable 5.1 on most work while costing 40 per cent less to run than Opus 5 on typical workloads.
A day earlier, investors had rewarded Meta for something different. Muse, its new personal AI agent, had overtaken ChatGPT to become the most downloaded free iPhone app in the United States less than two weeks after launch. Meta’s shares rose more than 11 per cent that day, helped by Muse’s early success and a Wells Fargo upgrade. By 21 September, Sensor Tower estimated that Muse had been downloaded more than 2.5 million times.
The contrast is revealing. Two frontier labs were competing on the price of intelligence. Meta was winning attention with a model that was not the most capable in the market.
That does not mean models have ceased to matter. It means that once a model becomes good enough and cheap enough, competition moves downstream—to distribution, product design, execution and control of the customer relationship.
The Price of “Good Enough” Is Collapsing
AI models have been getting cheaper for years, but the speed is easy to underestimate.
Epoch AI estimates that, since 2023, the cheapest cost of achieving a given level of benchmark performance has fallen by roughly 47 per cent per quarter—equivalent to an annual price-performance gain of about 13×. The researchers are careful about the limitations: benchmarks are not the same as useful work, the data cover only a few years, and real users do not constantly switch to whichever model happens to be cheapest. Even so, the direction is unmistakable.
The decline is not simply the result of better chips. More efficient architectures, caching, quantisation, distillation, mixture-of-experts designs and better inference software all allow providers to deliver similar capabilities with less compute. One recent analysis estimates that, after adjusting for competition and hardware improvements, algorithmic efficiency alone may be improving at roughly 3× a year.
But there is an important distinction between the price of a fixed level of capability and the price of the absolute frontier.
The former is collapsing. The latter remains expensive. The same research finds that the cost of running frontier models has increased by a factor of roughly 3 to 18 each year as models have grown and reasoning traces have become longer. In other words, yesterday’s exceptional capability quickly becomes affordable, while today’s very best capability continues to command a premium.
So the model layer is not disappearing as a source of advantage. Frontier capability still requires enormous capital, specialised talent and infrastructure. What is changing is the size of the “capability overhang”: more and more products can now be built with models that are not the best, only good enough.
Muse shows what happens when that overhang reaches the consumer market.
From Answers to Outcomes
A conventional chatbot waits for a question and returns an answer. Muse is designed to take a goal, break it into steps and keep working until it reaches a result.
According to Meta’s launch announcement, every user gets a dedicated cloud-based virtual machine with its own browser. Muse can open websites, fill in forms, connect to outside services and continue working after the user closes the app. It can be used through a standalone app or directly inside WhatsApp, and it pauses for approval before sensitive actions such as sending an email or making a purchase.
Early users have reported striking results. In one community-submitted example, Muse generated three options for a Santorini trip in thirteen minutes, with a best-value itinerary of $6,695; nothing was booked without approval. A founder said Muse replaced his car insurance in about five minutes, saving him $3,500 a year. Another community report said it sorted more than 40,000 emails and called Costco customer service, returning a transcript. These examples are self-reported rather than independently tested, but they illustrate the product’s promise: not better conversation, but delegated execution.
The reality is still uneven. PYMNTS tested Muse on three ordinary errands—reordering toilet paper from Amazon, buying a Domino’s pizza and booking a restaurant through Resy. It failed to complete all three. The obstacle was not reasoning so much as the messy infrastructure around the task: account permissions, separate authorisation flows and checkout.
That gap matters. An agent only saves time if delegation is easier than doing the job yourself. Today, Muse can look less like an autonomous assistant than a capable intern who still needs access, instructions and supervision.
Yet even its failures reveal why the category matters. The bottleneck is moving beyond the model. It now includes identity, permissions, payments, merchant access and the design of the hand-off between human and machine.
Meta’s Real Advantage Is Reach
Why did Muse break through when OpenAI has Codex and Anthropic has Claude Code—agents that are already useful and, in many tasks, more capable?
Part of the answer is the audience. Codex and Claude Code began with professionals: software developers, analysts and other knowledge workers willing to learn a new tool. Meta can start with the mass market.
Across Facebook, Instagram, Messenger and WhatsApp, 3.6 billion people used at least one Meta app each day in June 2026. Because Muse is available inside WhatsApp, many users do not need to discover a new product, create another account or learn an unfamiliar interface. They can message an agent in the same way they message a person.
This is why Ben Thompson’s observation is so important. Writing in Stratechery, he noted that “Muse Spark 1.3, the most advanced Meta model, is still not state-of-the-art”. But it is good enough to support what he called the best and most approachable personal agent he had tried.
That is a warning for the frontier labs. A superior model does not automatically produce a superior consumer business. Once a model clears the threshold required by the product, distribution and habit can matter more than a few extra benchmark points.
A personal agent may also be stickier than a chatbot. Switching chatbots is easy. Switching an agent that knows your preferences, has access to your services and has become embedded in your routines is harder. The durable asset may not be the model itself, but the context, permissions and workflows accumulated around it.
The Agent Wants the Customer Journey
The clearest sign of the coming conflict is the opposite reaction from Amazon and Shopify.
Amazon has blocked Muse from shopping on its website. Publicly, it has raised legitimate concerns about authorisation, credentials, transparency and whether third-party agents should be allowed to enter a service without permission. But the strategic stakes are just as important. Amazon owns a valuable front end: search results, recommendations, sponsored listings and more than $68 billion in annual advertising revenue. If an outside agent handles discovery, comparison and checkout, Amazon risks losing the customer’s attention even when it still fulfils the order.
Shopify has taken the opposite approach. It is working with Meta to make Shop Pay available to Muse across Shopify-powered stores. That makes sense. Shopify earns money by helping merchants run stores and process transactions; it does not need to preserve a central consumer homepage. If agents bring independent merchants incremental sales, Shopify can benefit without owning the browsing experience.
The split reveals what agents threaten to become: a new layer between demand and supply.
Recommendation feeds are built to keep people engaged. Their business value rises with attention and time spent. Agents promise the opposite. Their value lies in reducing the time between intention and outcome. A feed wants you to keep scrolling; an agent wants to make the decision, complete the task and return your time.
That difference could reshape digital commerce. If the agent decides which products to compare, which merchant to trust and when to buy, the agent—not the marketplace or brand—owns the point of decision.
Meta’s Business-Model Paradox
Mark Zuckerberg has said that Muse should eventually pay for itself by helping users save or earn money, with Meta taking “a very small cut” of transactions. As he put it: “We think that this thing is actually going to make you money and save you money, and that is how it’s going to pay for itself.”
It is an attractive proposition, but it contains a structural tension.
First, agents have real marginal costs. Each user needs cloud computing, inference, browser rendering, security monitoring and retries when a task fails. Meta can subsidise those costs from its advertising business in a way that most start-ups cannot, but the costs do not disappear.
Second, helping consumers save money often means compressing merchant margins. The more effectively an agent compares prices and negotiates, the less room a merchant has to pay commission.
Third, an agent-mediated order can weaken the merchant’s relationship with the customer. The merchant may receive a name, an address and a payment, but lose the browsing, storytelling and repeated exposure that build brand preference. On the next purchase, the agent can run the comparison again. Winning one order may create little advantage in winning the next.
For merchants, that begins to look less like a customer relationship and more like a series of one-off auctions. For the agent platform, meanwhile, the economics must cover a costly front end with a commission extracted from sellers whose margins it is simultaneously trying to reduce.
That does not make the model impossible. Subscriptions, merchant tools, preferred integrations and payment services could all contribute revenue. But it explains why agentic commerce is not simply a better version of online shopping. It rearranges who owns demand—and therefore who captures the profit.
Meta Has Been Here Before
In 2016, Zuckerberg used Facebook’s F8 conference to introduce bots for Messenger. The app had just passed 900 million monthly users, and more than forty partners had built commerce, transport and customer-service experiences on the platform.
“I have never met anyone who likes calling a business,” he said, before demonstrating how a user could message 1-800-Flowers instead. The promise was familiar: no new app, no phone call, just send a message and get something done.
It did not become a new operating system for commerce. Messenger bots largely settled into customer service.
A decade later, Meta is returning to almost the same interface with a far more capable technology. The difference is that Muse is not merely a script owned by a business. It is intended to be an agent working for the user, capable of navigating across businesses and services. The models are better, the distribution is broader, and the infrastructure for identity and payments is more mature.
Whether that is enough remains unclear. Reliability is inconsistent. Trust is fragile. Merchants may resist an intermediary that takes the customer relationship, and the economics of providing everyone with a cloud computer are still unproven.
But Muse has already clarified the direction of travel.
The AI race has not moved on from models; it has widened beyond them. As capable intelligence becomes cheaper, competitive advantage shifts towards whoever controls the interface, the user context, the workflow and the transaction.
The best model may win the benchmark. A good-enough model, placed inside the right product and pushed through the right network, may win the market.