Recently, Martin Casado, a partner at a16z, made a judgment that makes model providers somewhat uncomfortable: as compute supply is unleashed and open-source models mature, the value in the AI industry chain will be redistributed. The application layer will capture an increasingly larger share, while the profit margin of model providers will be eroded. The underlying logic is straightforward: as models become commoditized, they become homogeneous and interchangeable, making it increasingly difficult to differentiate based on the models themselves.
This judgment directly refutes the mainstream narrative of the past two years: models will consume applications. Over the past two years, the industry has reached a high consensus that if a model is strong enough, it will consume applications layer by layer. Applications are merely a thin shell for the models, ready to be swallowed by model providers at any time. This narrative is popular because model providers sit at the very top of the value chain, holding both pricing power and profits. For example, in applications such as coding and video, dominant products all come from dominant model providers, where model capability is almost equivalent to application capability.
Now, the direction of this value chain is reversing. Especially in office applications, the dominant product WorkBuddy does not come from the strongest model provider, and office applications are not deeply bound to a single model, having opened up to multi-model access. It can be seen from this that the strength of a model no longer determines the superiority of a product, and the voice of the model within applications is declining. iFenxi believes this is a watershed moment for the decision-making power of the value chain.
Model commoditization allows models to be swapped at any time, but task data will only accumulate on the application side. Consequently, the decision-making power of the value chain shifts from model providers to application providers. The driving force behind this is exactly the Harness, which has been hotly debated this year. Harness refers to the system outside the model that translates model capabilities into specific task executions, encompassing the Agent Loop, context management and memory, tool calling, task orchestration, and result feedback. It cannot increase the intelligence ceiling of the application, but it determines how intelligence is correctly utilized.
01. Determining the Success or Failure of Tasks: Shifting from Models to Harness
In August this year, OpenAI disclosed a set of data showing that its then-latest model, GPT-5.6 Sol, through Harness adjustments, increased its score on the ARC-AGI-3 benchmark from 13.3% to 38.3%.
Figure 2: Score performance of GPT-5.6 Sol before and after Harness adjustments
From the evaluation results, the model itself has not changed, yet the score has increased by nearly three times. This is because the Harness has unlocked the capabilities the model already possessed but had not previously released. This is precisely the value of the Harness layer.
Besides OpenAI, multiple research papers in 2026 have also conducted relevant controlled experiments to quantify the value of Harness. For example, in a paper by scholars from Tulane University and others ("Stop Comparing LLM Agents Without Disclosing the Harness"), the experiment selected the same batch of long-horizon tasks. First, the Harness was fixed while different models were swapped and tested; then, the model was fixed while different Harnesses were swapped and tested. The fluctuation range of the two sets of results was compared. The larger the fluctuation value, the greater the impact of this variable on the result. The experimental result showed that the performance fluctuation caused by changing the Harness was 7.8 times that of changing the model. The role of Harness is reflected not only in task execution effectiveness but also in cost control.
In the aforementioned GPT-5.6 Sol test, the token consumption under the Harness framework was 1/6 of that without the framework. In addition, the AI company Writer also conducted another experiment ("The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI"). For the same batch of long-horizon tasks, only the Harness was changed while the model remained unchanged. As a result, the cost per task decreased by 41%, the time consumed decreased by 44%, and token consumption decreased by 38%, while the quality remained the same. In other words, with the same model being scheduled by a more rational Harness, the resource consumption for completing tasks can be reduced by nearly half.
02. Static Harness Only Provides a Temporary Lead; Evolving Harness is the True Barrier
The aforementioned static Harness engineering can be easily replicated, and model providers will not be cannibalized by application providers as a result. iFenxi believes that the data-driven self-evolving closed loop of the Harness is the true barrier, and this closed loop can only be built on the application side. This is also the core reason why the AI narrative is shifting from models to applications.
A real long-horizon task consists of a series of calls: first understanding the requirements, then breaking down the steps, calling tools, checking results, retrying upon errors, and finally summarizing. Only the application that initiates the task knows the causal chain between these calls. What model providers see are merely several received requests; they neither know whether these requests belong to the same task nor know the sequence and success or failure among them.
Therefore, model providers cannot refine the Harness for real tasks. For the application side, however, the situation is completely different. For example, every time an office product integrates a new industry, it adds a new type of task pattern and a new portion of real data.
The more industries integrated, the more comprehensive the scenarios the Harness has seen, and the more accurately it handles subsequent new tasks. Moreover, the marginal cost of integrating new industries also decreases accordingly. Therefore, for application providers, this is a positive cycle: the stronger the Harness system, the more capable it is of handling more complex scenarios, which has also spawned a batch of new opportunities for Harness self-evolution. This self-evolution has a ceiling; it can make the model used more correctly and better, but it cannot change the capability ceiling of the model itself.
03. Countermeasures of Model Providers: Open-Sourcing Harness
To counter this change in the value chain, model providers use open-sourcing Harness as a means to try to dismantle the barriers built by application providers. In the second half of the year, domestic players such as DeepSeek and Qwen have successively open-sourced their respective Harness or Agent frameworks, allowing developers to quickly start on a ready-made foundation without building from scratch.
The intention of model providers is very clear: when Harness becomes a publicly available foundational component, all parties in the ecosystem can optimize the Harness for their respective application scenarios, and the space for application providers to build advantages through in-house development is consequently compressed. Open-sourcing by model providers kills two birds with one stone: it breaks the competitive barriers of opponents and attracts more developers to their side, anchoring the ecosystem around the model layer.
Table 1: Harness Frameworks Open-Sourced by Model Providers
iFenxi believes that model providers open-sourcing Harness frameworks will temporarily impact application providers that develop their own Harness. Application providers will never open-source their own Harness, because it is the entry point for connecting with users and accumulating data, as well as the accumulation of know-how gained through years of trial and error. Once open-sourced, it is equivalent to making their playbook public and weakening their commercialization capabilities.
But this impact is temporary. The open-sourcing by model providers can only impact the static Harness barriers of application providers. When self-evolution capability becomes the main body of the Harness, this impact will dissipate accordingly. Open-source versus closed-source will no longer be the key variable; the data closed loop will be the true barrier. Model providers represent the open-source ecosystem, while application providers represent the closed-source ecosystem. In the future, whichever ecosystem the data closed loop resides in will have the stronger voice.
04. The Boundary of Cannibalization: Applications Take the Entry Point, Models Retain the Ceiling
The cannibalization of models by applications also has its boundaries. The endgame will not conclude with applications consuming models, nor does it mean that applications can return to the old path of self-training models.
First, what application providers cannibalize is the model's mindshare among users. In the past, model capability equaled application capability, and users had strong brand awareness of models; now, model capability and application capability are gradually decoupling. Users' mindshare is gradually being captured by application brands, while model providers retreat behind the scenes, becoming replaceable suppliers.
Second, the model barrier still exists. As mentioned earlier, the ceiling of task execution capability is determined by the model, and the Harness improves the stability, cost, and success rate of tasks within the model's capability range. In other words, whether ceiling-level, highly difficult tasks can be solved depends on the model.
Li Qiang, Vice President of Tencent, gave an accurate description of the boundary between Harness and models: "The model is the engine, and the Harness is the engineering that turns the engine into a complete vehicle. The engine determines the upper limit, while the Harness determines whether it can run, how far it can run, and how steadily it runs." In other words, the model represents the upper limit, and the Harness determines how close it can get to that limit.
Combining the above two points, iFenxi believes that the cycle of applications cannibalizing models will be very long, and enterprises and developers will not currently place the importance of Harness above that of models. On the one hand, neither the Harness framework nor the models have reached a steady state. If the quality of application output changes, it is difficult to correctly attribute the cause, leading to a lack of direction in application iteration. On the other hand, the value of breaking through the task ceiling remains very high, and in this regard, they still have to rely on SOTA models.
In the future, the relationship between applications and models will gradually shift from competition to division of labor and collaboration. Model providers retreat to defend the capability ceiling, while application providers occupy execution and the entry point. The value of both parties is refocused: model providers focus their energy on continuously pushing up the model capabilities, while application providers focus their energy on running real tasks through smoothly. Both sides transition from a competitive relationship in the upstream and downstream to a relationship of division of labor and complementarity.