Analysts note that the capabilities of current state-of-the-art models have become "excessive," and model performance is no longer the most important competitive advantage for AI companies; the focus of competition is shifting from the model layer to the application layer.

Author: Long Yue

Source: Wall Street Journal

On September 21, Meta rose over 11%, posting its largest single-day gain since April last year, with its market capitalization increasing by more than $190 billion in one day.

The market catalyst was Meta's recently launched personal agent, Muse. According to Sensor Tower data, the app surpassed 902,000 cumulative downloads on U.S. iOS within six days of launch, significantly outperforming Meta’s previous product. By Monday, Muse had topped both the U.S. iOS and Google Play free charts, displacing ChatGPT.

The market is paying for the hype around Muse. Technology analyst Ben Thompson commented in his article “Frontier Overhangs”: “Muse Spark 1.3 is not the most advanced model, but that’s precisely the warning sign—a model that isn’t leading the pack is already sufficient to power a highly addictive personal agent product.”

Thompson believes that the capabilities of current cutting-edge models have become "excessive," and model capability is no longer the most important moat for AI companies; the next competitive frontier may be the "application moment."

After models surpass the tipping point, the focus of AI competition is shifting. From consumer-facing products like Meta’s Muse to enterprise solutions like Tencent’s WorkBuddy, a common signal is emerging: what increasingly determines a product’s value is not the model itself, but control over harnessing, workflows, and user access points.

On September 22, Tencent followed Meta's overnight rally, surging over 6% in early trading in Hong Kong, reaching a intraday high of HK$463.40.

Meanwhile, research from consulting firms and venture capital firms is validating the same conclusion from multiple perspectives. Gartner warns that by 2030, up to $234 billion in enterprise software spending is at risk from “Agent arbitrage”; a16z explicitly advises AI application companies to move away from token-based pricing and adopt outcome-based pricing; and data from ICONIQ’s survey of over 300 AI software companies shows that the application layer has become the industry’s primary focus, with AI product gross margins expected to rise from 45% in 2025 to an estimated 53% in 2026.

Is the “application moment” approaching?

Thompson introduced a key concept: when the capabilities of frontier models continue to advance, but the adoption of these capabilities in real-world products and user scenarios lags behind, a "frontier overhang"—an excess of model capability—emerges.

He cited a specific example: the AI narrative gaming company Fable launched a new-generation model, but market demand remained lukewarm. This was not due to poor product quality, but because the capabilities of existing models were already “sufficient” for this use case—stronger models did not deliver perceptible improvements in user experience, and thus failed to translate into higher willingness to pay or greater market share.

Thompson believes that when performance is not yet sufficient, the one who tightly couples the model with the application layer (harness) wins; once the threshold of "sufficient" is crossed, customers begin to care about time-to-market, convenience, customization, and data handling—这时 modularity gains an advantage.

This detail is indicative. It shows that, in certain scenarios, the marginal value of model capabilities is diminishing. When a "stronger model" no longer equates to a "better product experience," the competitive advantage at the model layer begins to decouple from the business value at the application layer.

Thompson further pointed out that, in this context, what truly matters is the "harness"—how to effectively organize, schedule, and encapsulate the model’s capabilities to reliably and stably accomplish tasks in specific scenarios. The model is the engine; the harness is the transmission system. Without the latter, no matter how powerful the engine, its output cannot be effectively delivered. This framework provides a clear coordinate system for understanding the product logic behind Meta Muse and Tencent WorkBuddy.

C-end: Meta Muse's product bet

Meta has launched Muse, a classic example of an "application-layer bet." According to Meta's official description, Muse is positioned as a personal AI agent powered by the underlying Muse Spark 1.3 model. However, the product’s key selling point is not its model size or benchmark scores, but its ability to actually accomplish tasks—helping users plan, execute steps, and track progress by genuinely “getting things done” in everyday life scenarios.

Behind this positioning is Meta’s assessment of the AI product competitive landscape: at a time when model capabilities are relatively abundant, what users truly lack is not “smarter models,” but “tools that can get things done for me.” Muse aims to become the user’s personal execution layer, not just a question-and-answer interface.

Looking further, Muse’s product focus has shifted the breakthrough for consumer-facing agents from “whether the model is smart enough” to “whether users can confidently delegate tasks and whether the product can connect to real-world actions.” The Muse Spark series of models has helped Meta fill in foundational capabilities, but what determines whether an agent can enter users’ daily lives goes beyond reasoning and generation—it also includes authorization, trust, and execution workflows. To this end, Muse has integrated independent execution environments, credential protection, and critical action approvals into its product design to establish clear authorization boundaries. At the same time, by combining long-term memory with background execution capabilities, Muse enables agents to continuously advance tasks around user goals, rather than stopping at one-time Q&A or command responses.

From a business logic perspective, this design has profound intent. Once a personal agent becomes deeply integrated into a user’s daily workflow, it accumulates vast amounts of personal data, preferences, and usage patterns, creating user loyalty that is difficult to migrate. This loyalty is the true moat—not the capability of any specific model version.

The market's positive response to Muse's release partially validates this logic: investors' interest in "usable" Agent products has surpassed their focus on "more powerful" models.

B2B: Tencent WorkBuddy's Enterprise Path

On the enterprise side, Tencent WorkBuddy follows a similar product logic. According to Tencent Cloud’s official description, the core strength of WorkBuddy Enterprise lies not in providing a generic chat interface, but in connecting various internal systems and workflows, enabling cross-platform task execution, and truly embedding AI capabilities into daily business operations.

This represents a critical leap for enterprise AI from a “tool” to the “execution layer.” In the past, companies purchased AI products essentially as smarter search or Q&A tools; WorkBuddy points toward a future where AI becomes an Agent capable of proactively initiating, executing, and completing tasks, directly participating in business processes.

The business implications of this shift are equally profound. As AI becomes capable of completing tasks across multiple software systems, the "functional value" of traditional enterprise software faces direct challenges. In its July 2026 report, Gartner explicitly stated that the rise of agentic AI puts as much as $234 billion in enterprise application software spending at risk—because agents can bypass or replace workflows that previously required multiple standalone software solutions, potentially eroding the traditional software "moat" of functionality.

For enterprise software companies, this is a structural pressure that must be taken seriously.

Repricing Business Value: From Tokens to Outcomes

The "sufficiency" of model capabilities is driving a fundamental shift in AI business models.

Venture capital firm a16z explicitly highlighted this trend in its report, "You are not a model. Don't price per token.": The true value of AI applications comes from the data, tools, workflows they integrate, and the measurable outcomes they deliver—not from which model is invoked or how many tokens are consumed. In other words, the pricing logic of "selling model capabilities" is being replaced by the logic of "selling results."

a16z outlines three pricing models: charging for model access on a per-token basis; transforming the model into useful work and pricing it according to value units recognizable to customers, typically points; and delivering clear, attributable business outcomes with pricing based on results. Among the 50 enterprise AI technology buyers surveyed, 27 preferred points tied to recognizable workloads, while 14 preferred pay-per-token pricing.

This has a direct impact on the value chain distribution in the AI industry. If the model itself gradually becomes a foundational capability—similar to computing power in the early days of cloud computing—then the companies that will truly capture commercial value are those at the application layer that can build specific products, vertical workflows, and verifiable ROI on top of the model.

McKinsey’s report, “Where AI Agents Pay Off,” provides corporate validation: as token costs continue to fall, businesses evaluating AI investments have shifted their focus from “How powerful is this model?” to “How much measurable efficiency gain and economic return can this agent deliver?” Verifiable ROI is becoming the core criterion for corporate AI procurement decisions.

This means that the question AI application companies need to answer has shifted from “What model are we using?” to “How much money can we save our customers and how much efficiency can we improve?”

Industry Landscape: Application-Layer Competition in the Multi-Model Era

ICONIQ surveyed over 300 AI software companies in its 2026 State of AI Report, revealing a clear industry trend: nearly two-thirds of surveyed companies are developing horizontal or vertical AI applications, and the use of multiple models has become standard.

The fact that "using multiple models has become the norm" is itself strong evidence of the commoditization trend of models. When enterprises can flexibly switch between different models based on task requirements, the bargaining power of model providers is diminished, while application-layer platforms that can effectively integrate multiple models and build stable workflows gain stronger bargaining power.

ICONIQ's report also tracks changes in AI revenue and gross profit margin. The importance of this metric lies in its ability to help the market determine whether commercial value capture at the application layer is genuinely occurring—or remains merely narrative. For investors, this is a key indicator for distinguishing between “AI application stories” and “AI application businesses.”

From an industrial structure perspective, the emerging landscape is characterized by the model layer (OpenAI, Anthropic, Google, Meta, etc.) providing foundational capabilities, the application layer (various vertical agents and workflow platforms) transforming these capabilities into deliverable outcomes, and traditional enterprise software companies facing pressure to be bypassed or replaced. The distribution of value among these three layers will be the most significant dynamic to monitor in the AI industry over the coming years.

The traditional SaaS seat logic is being disrupted—if agents can complete tasks across systems, employees no longer need to frequently switch between every traditional software interface. The software is still running, but the interface may retreat to the background.

Gartner predicts that by 2030, up to $234 billion in enterprise application software spending will be at risk due to Agentic AI, accounting for approximately 20% of enterprise application SaaS spending.

George Brocklehurst, Managing Partner at Gartner, said: “Agentic AI changes the economic logic of software. Agent systems deliver results directly, bypassing applications that heavily rely on user interfaces, making software invisible.” The traditional link between the number of agents and software revenue growth may therefore be weakened.

The pressure first falls on vendors that rely on dashboards, feature modules, and agent licensing fees. Enterprise buyers won’t increase their budgets just because an AI button has been added; they will ask whether the agent reduces manual operations, shortens processes, improves conversion rates, or lowers unit costs.

But this is not the end of traditional SaaS. Vendors with access to industry data, customer relationships, system integrations, and control over business processes can still reposition themselves as the agent execution layer. The risk lies in charging for access to the interface; the opportunity lies in owning the task execution and delivery of results.

Is it possible to monetize directly with a leading model?

It should be clear that the above analysis does not lead to the conclusion that "models are unimportant." The capabilities of cutting-edge models remain the foundation of the entire AI industry; without sufficiently powerful models, building applications would be impossible. The real question worth discussing is: Can leading-edge models still be directly translated into product advantages and commercial value?

Current evidence suggests that this conversion path is becoming longer and more complex. The poor reception of Fable’s new model illustrates that, in certain scenarios, stronger models no longer deliver perceptible improvements in user value. The product strategies of Meta Muse and Tencent WorkBuddy show that leading tech companies are building differentiation through “application layer + workflow + user entry points,” rather than relying solely on model capabilities. Research from a16z and McKinsey indicates that enterprise customers’ payment logic is shifting from “model capabilities” to “quantifiable outcomes.”

This means competition in the AI industry is entering a new phase: model capabilities are a necessary condition, but no longer sufficient. Those who can transform general-purpose models into concrete product experiences, stable workflows, and verifiable business returns will gain a true competitive advantage in the next stage.

This shift suggests a valuation framework worth reexamining: in the AI field, while model parameter scale and benchmark rankings remain important, depth of application-layer workflows, user entry point stickiness, and verifiable ROI may now be more predictive indicators of commercial value. As models become “good enough,” the battlefield for the next competition has quietly shifted.