Summary of Key Points
The AI industry is experiencing a counterintuitive price reversal: proprietary models (such as OpenAI and Anthropic) are becoming cheaper, while open-source models are getting more expensive. This is not simply a matter of "free vs paid" but is the result of a complex interplay of factors including cost structures, business models, and geopolitics. Proprietary models can drive down prices through competition and economies of scale, whereas the hidden costs associated with open-source models (such as computing power and maintenance) increase as their usage expands. At the same time, the business models of both types are beginning to merge—proprietary models offer "convenience services," while open-source models rely on cloud services for revenue generation. The definition of open-source is also evolving, shifting from being completely transparent to semi-proprietary, and geopolitics is starting to influence model selection. The ultimate trend is a coexistence of both: companies will use proprietary models for critical security tasks and open-source models for high-frequency, less sensitive tasks. Model companies that cannot provide comprehensive delivery services will be marginalized.
1. Counterintuitive! Proprietary Models Are Getting Cheaper, While Open-Source Models Are Rising in Price?
Do you think using open-source models for free means saving money? Wrong!
- Reasons for the decrease in proprietary model prices: The three major proprietary giants (OpenAI, Google Gemini, and Anthropic) are competing on price; the more people use their models, the lower the cost per token. Additionally, technological improvements have reduced costs (for example, Anthropic increased its gross margin from -94% to 60%). As a result, the API prices for proprietary models have dropped by 88% over three years, with the current median price for cutting-edge models being only $4.5 per million tokens and $1.93 for mid-range models.
- The truth behind the increase in open-source model prices: While the weights used in open-source models are free, running them incurs additional costs—renting GPUs, hiring engineers to tune parameters, fixing bugs, dealing with memory overflow issues, etc. These hidden costs rise significantly with increased usage. For instance, the token share of Chinese-made open-source models has risen from 2% to 46%, leading to higher total costs. Bloomberg data shows that the "expenditure-weighted price" (the actual amount users pay) for open-source models has nearly matched that of proprietary models.
In short, open-source models are seemingly free but require hidden payments, while proprietary models have clear pricing but are becoming cheaper.
2. The Logic of Making Money Has Changed: Proprietary Models Offer Convenience, Open-Source Models Rely on Additional Services
AI companies are no longer competing on the quality of their models but on the delivery services they provide:
- New business models for proprietary models: Anthropic’s annual revenue has soared from $9 billion to $60 billion, not because its models are significantly better (the performance difference is now only 3%), but because it offers additional services such as liability protection, service level agreements, data compliance, and enterprise-level access controls. Companies are willing to pay more for the convenience of not having to manage these aspects themselves.
- Open-source models’ strategy: Alibaba’s Qianwen platform offers a fully open-source framework, but users must use its cloud services for inference. Other companies like DeepSeek and ByteDance combine open-source foundations with proprietary advanced models and additional services. For example, the highly parameterized Kimi K3 (2.8 trillion parameters) is priced at $100 per million tokens, generating revenue through cloud services.
In essence, the value of an AI model lies in the ability to deliver it reliably to customers.
3. The More You Save, the More You Spend? The AI Version of the Jevons Paradox
This phenomenon is similar to the 19th-century steam engine paradox: more efficient engines led to increased coal consumption because more people could afford them. In AI, the same pattern is observed:
- Inference costs have decreased by 50 times (from $20 to $0.4 per million tokens), making AI more accessible to businesses. A MIT study shows that switching to open-source models can save 70% of costs, but the saved money is often reinvested in more AI applications, leading to an overall increase in model usage and thus in computing power consumption.
- Morgan Stanley predicts that as end-user costs decrease, the total market will expand. Funds flow from model development to cloud providers and internal company expenses (such as deployment and maintenance), putting significant pressure on engineers.
4. Is Open Source Really “Open” Anymore? Maybe It’s Semi-Proprietary
Traditional open-source models provided all training data, code, and weights. Nowadays, many only offer the weights while keeping the training processes secret:
- According to MIT and HuggingFace, 79% of open-source models were truly open-source in 2022, but this figure is down to 39% in 2025. The trend is for companies to only share the weights, not the training data. For example, you can use an open-source model but not know how it was trained or modify its core logic.
This shift allows companies that control the training data to maintain pricing power. After the release of Kimi K3, OpenAI’s valuation dropped by $314 billion, as revenue went to the company that owns the training data.
5. Geopolitics Are Entering the Picture: AI Model Selection Has Become a Compliance Issue
AI is no longer just a technical matter; geopolitics are playing a role:
- U.S. senators have called for banning government use of Chinese-made open-source models, but the White House has exempted some from security tests. Over 200 Silicon Valley companies oppose this ban, and DoorDash has been investigated for using such models.
- Chinese-made open-source models account for more than 30% of usage on OpenRouter, highlighting the importance of compliance in model selection.
6. The Final Outcome: A Coexistence of Both Models
Morgan Stanley predicts a hybrid approach where companies will use proprietary models for critical tasks (e.g., financial risk management) and open-source models for less sensitive, high-frequency tasks (e.g., customer service responses).
Companies that only offer models without comprehensive delivery services (compliance and maintenance) will be at a disadvantage. The future belongs to those who can address real business challenges and provide the necessary solutions.
In summary, the AI industry is shifting from a focus on model performance to a focus on providing practical solutions that meet companies’ needs—convenience, compliance, and stability. Those who can do this will succeed in the market.