Summary of Key Points
OpenAI and DeepSeek both open-sourced their Agent frameworks (Harness) within a week, but the nature of their open-source initiatives, business models, and goals differ significantly: OpenAI released a fully developed task execution capability (Codex Harness) with the aim of enabling more products to generate revenue through its models via APIs; DeepSeek, on the other hand, provided component-based assembly rules (DSH) in an effort to establish its models as the standard in the ecosystem. This open-source movement is not a triumph of technical ideals but rather a commercial battle centered around models and ecosystems. Both companies are focusing on “opening up the periphery while securing the core,” shifting their strategic barriers to more valuable areas.
1. Both are Open-Sourcing Harness, but They Are Playing Different Games
First, let’s understand what Harness is: Large models are like trains (capable of reasoning), but to make them perform practical tasks (such as financial month-end calculations or code reviews), a “track” is needed—something that connects the model, tools, and user systems. Harness is responsible for breaking down tasks, invoking tools, tracking progress, and even pausing sensitive operations awaiting user approval, transforming the model from a “question-answerer” into a “continuously working agent.”
However, the open-sourced Harnesses from these two companies differ greatly:
- OpenAI’s Codex Harness: It’s like a pre-made robot with a fixed form and control logic. You can equip it with different “tools” (such as financial software interfaces or code review plugins) to perform various tasks, but the core “brain” (the Codex model) remains OpenAI’s. Developers don’t need to write complex logic; they can simply adapt it to industry-specific contexts. The final product may bear their own brand, but OpenAI still charges for using its model.
- DeepSeek’s DSH: It’s like a robot assembly platform. You can choose different “brains” (models), “arms” (tools), and “memories” (storage) and assemble them according to its interface rules. For example, for a financial agent, you can use a cheaper model or upgrade to a more secure sandbox without having to rewrite everything. DeepSeek aims to make these rules the industry standard, ensuring that even if models change, its standards will continue to attract developers, making its models more attractive due to their cost-effectiveness.
2. Has OpenAI Changed Its Strategy? Because What It Opens Source Doesn’t Affect Its Profitability
Just a month ago, OpenAI executives criticized Chinese companies for open-sourcing models as a form of “AI communism.” Now, however, OpenAI is open-sourcing Harness. This isn’t a contradiction; it reflects a shift in business strategy:
- Model Open-Sourcing Is Not an Option: Models are OpenAI’s main source of profit (billions were spent on training them, and revenue comes from APIs and subscriptions). If models were open-sourced, customers would deploy them themselves or opt for cheaper alternatives, depriving OpenAI of revenue.
- Harness Open-Sourcing Is an Option: Harness serves as a channel to sell models. By making it available, OpenAI allows more developers to integrate its Codex models into their products (such as financial software or operations tools), entering markets that it couldn’t reach on its own. Although Harness is free, OpenAI can still charge for API usage when these products utilize its models, effectively using a free channel to generate long-term revenue.
In other words, OpenAI is keeping its most profitable models while offloading the task of finding customers for them to developers, reaping the benefits without direct effort.
3. This Open-Sourcing Movement Is an Ecological Battle to Weaken Each Other’s Position
The simultaneous open-sourcing of Harnesses by both companies is not coincidental but part of a strategic attack:
- OpenAI’s Counterattack: With DeepSeek’s DSH, developers can freely switch models, which could lead to customers leaving OpenAI’s platform. Therefore, OpenAI is releasing Codex Harness to retain developers with its convenient, ready-to-use solution. Many companies are more interested in quickly launching products and won’t invest time in custom assembly, allowing OpenAI to maintain its ecosystem.
- DeepSeek’s Offensive Strategy: DSH makes models replaceable components, potentially turning OpenAI from a platform leader into a model supplier. For example, companies using DSH to build agents can choose between OpenAI’s GPT and DeepSeek’s models (if the latter is cheaper). This forces OpenAI to compete on model cost-effectiveness, which is DeepSeek’s strength (lower costs and better performance).
4. Reactions from Other Players
The reactions of other players in this battle are interesting:
- Anthropic: Remaining inactive for now. It generates revenue from Claude’s API and subscriptions and believes that a complete product is more important than open-sourcing the underlying technology. However, if OpenAI and DeepSeek’s Harnesses attract too many developers, Anthropic might need to open-source its own framework as well.
- Google: Has already open-sourced a similar framework (ADK) but without much attention. Google’s core business is cloud services, so it doesn’t urgently need to control agent assembly; it can still make money regardless of the models or Harnesses used by developers, as long as they run on its platform.
5. Open-Sourcing Doesn’t Remove Barriers; It Just Moves Them
This open-source movement isn’t about “breaking monopolies” but about “redrawing the boundaries of the market.” Companies open-source the parts that can expand the market while securing the core areas that generate profits:
- OpenAI focuses on protecting its models and using Harness as a channel to sell them.
- DeepSeek focuses on establishing standards and using its components as an entry point into the ecosystem.
- NVIDIA focuses on GPUs and CUDA, making its hardware more accessible through open-source tools.
In summary, this open-source movement isn’t about eliminating barriers but about shifting them to more valuable areas. The winner will be the one that can convince the entire industry to use its open-source components while still maintaining its proprietary core technologies.
In One Sentence: Both companies’ open-sourcing of Harnesses is aimed at boosting their core businesses. This battle is just beginning.