The article mainly discusses the challenges of billing models and engineering efficiency issues faced by artificial intelligence, especially large model services, in the Agent era. The article uses Anthropic’s decision to cut off subscription users of its Claude service from accessing third-party Agent frameworks (such as OpenClaw) as an introduction, discussing the deep-seated reasons, industry reactions, and possible future directions.
Luo Fuli, head of Xiaomi MiMo large model, expressed on X that Anthropic announced that Claude Pro and Max subscribers can no longer use their subscription quotas for third-party Agent frameworks such as OpenClaw, which forces heavy users to switch to pay-as-you-go APIs. Anthropic explained that the subscription pricing was designed based on the intensity of individual users, while the usage intensity of automated agent tools far exceeded expectations, resulting in the company bearing huge computational cost pressure.
Luo Fuli pointed out that Anthropic’s subscription system may be in a loss-making state. She analyzed that third-party frameworks such as OpenClaw have defects in context management, and a single request will trigger multiple rounds of low-value tool calls, which extremely wastes computing power, and its real API cost is often dozens of times the subscription price. She believes that this pain will force engineering progress, forcing framework developers to optimize context management and reduce invalid Token consumption.
At the same time, Luo Fuli warned large model companies not to blindly engage in price wars. She believes that selling Tokens at low prices but allowing third-party tools to exploit them is a trap, and low-quality Agent frameworks and unstable inference services will create a vicious cycle of user experience. She introduced Xiaomi’s MiMo Token Plan, emphasizing the pursuit of long-term stable delivery of high-quality services, and pointed out that the global computing power supply has not kept up with the demand of Agents, and the future lies in the coordinated evolution of “more Token-saving Agent frameworks” and “more efficient models.”
Regarding Luo Fuli’s views, the developer community reacted strongly, and the focus of discussion was on the structural rewriting of AI economics, the governance of computing power waste, and market elimination mechanisms. Developers believe that the orchestration layer is the core of the product, and the rough framework design has led to serious waste of computing power. In the future, the market will shift from “extensive burning of computing power” to “refined engineering architecture,” and clear Token quotas will become the key to fostering better product behavior.
[Machine Heart]
AI Token Economics Under Scrutiny: Implications for Crypto Investors
The recent controversy surrounding Anthropic’s restriction of Claude subscriptions for use with third-party Agent frameworks like OpenClaw reveals critical structural challenges in AI token economics that will significantly impact the crypto market. This isn’t merely a billing dispute but a fundamental reassessment of how value is measured and exchanged in the AI ecosystem.
Economic Mismatch and Market Realignment
Anthropic’s decision forces heavy users to shift from subscription-based to pay-as-you-go APIs, exposing a fundamental misalignment between current token models and actual usage patterns in an automated agent environment. Luo Fuli’s analysis that “real API cost is often dozens of times the subscription price” suggests that many AI crypto projects may face similar existential challenges if their token economies don’t account for automated consumption patterns.
This creates significant risks for tokens built on assumptions of human-level usage. Projects with inefficient architectures or opaque economic models could face margin erosion as compute costs rise and usage patterns evolve. The market is likely to undergo a painful realignment, with tokens demonstrating clear value propositions and sustainable economics emerging as winners.
Efficiency as a Competitive Moat
The developer community’s focus on “refined engineering architecture” over “extensive burning of computing power” signals a paradigm shift. For crypto investors, this creates a new investment thesis: tokens that demonstrably reduce waste in both token consumption and compute usage will develop competitive moats. Projects with context optimization, efficient inference, and clear token quotas will attract both users and capital.
Conversely, tokens associated with inefficient frameworks or engaged in unsustainable price wars without regard for long-term viability face significant downside risk. The article’s warning about “low-quality Agent frameworks and unstable inference services creating a vicious cycle” should serve as a red flag for investors.
Decentralized Compute Infrastructure Opportunity
The global compute supply-demand imbalance highlighted in the article presents a strategic opportunity for decentralized computing networks. Crypto projects offering alternative, potentially more efficient infrastructure solutions could position themselves as beneficiaries of the centralized providers’ growing pains. The market may reward protocols that offer transparent, measurable efficiency metrics and sustainable economic models.
Token Standardization and Economic Design
The industry’s shift toward “structural rewriting of AI economics” suggests we’re in early innings of token standardization for AI applications. Projects that can pioneer balanced economic models—combining subscription predictability with token flexibility—may emerge as leaders. Governance tokens allowing stakeholders to shape these economic parameters could see increased relevance as the market matures.
For investors, the key takeaway is that raw token consumption metrics alone no longer suffice. The market will increasingly reward tokens demonstrating efficient usage patterns, sustainable economics, and technical innovation that reduces waste. As Luo Fuli wisely notes, the future lies in the coordinated evolution of “more Token-saving Agent frameworks” and “more efficient models”—a lesson equally applicable to the crypto AI ecosystem.