Summary of Key Points
Recently, companies have encountered the issue of "ineconomical token usage" when employing AI programming tools: they spend a significant amount of money (in tokens) without obtaining corresponding value. Examples include Microsoft revoking the license for Claude Code and Uber using up its entire 2026 AI budget in just four months. The reasons for this include rising model prices, hidden inefficiencies in agent design, limited use cases, and uneven risks across the AI industry chain. To address this problem, efforts must be made to optimize technology (to reduce consumption) and expand use cases (to increase value), thereby bringing token usage back in line with a reasonable Return on Investment (ROI).
Detailed Analysis
1. Rising Model Costs Straining Corporate Budgets
AI models are like "intelligent service packages" that require payment for each use (in tokens). Recently, prices have been increasing for both high-end and general-purpose models:
- High-end Models: Hierarchical Pricing
Anthropic (the parent company of Claude Code) divides its models into three tiers: "Flagship (Opus)," "Mid-range (Sonnet)," and "Lightweight (Haiku)." The price of the high-end Opus model is five times that of the mid-range Sonnet, and it continues to rise. For example, the newly released Mythos Preview costs 1.6 times more than the previous version. Since companies using these models are less sensitive to price and place more emphasis on efficiency, Anthropic has generated substantial revenue (reaching $45 billion in 2026).
- General/Publicly Available Models: Apparently Inexpensive, but Actually Costlier
Models like GPT mini and Gemini Flash, although cheaper, have high usage costs due to their widespread adoption. For instance, the price of Haiku has increased by 20% annually, and the output cost of Gemini Flash has doubled. This is because these models handle most routine tasks (such as coding and document processing), leading to high consumption and unsustainable low prices.
- Consequence: Rising Costs for Enterprises
Model manufacturers' revenue has soared, but companies' token bills have become increasingly burdensome, contributing to the issue of "ineconomical token usage."
2. Hidden Wastage in AI Agents
Many tokens are wasted when AI agents perform tasks:
- Contextual Loops: Reusing Historical Data
AI systems incorporate previous conversations, code, and logs into their processing, leading to repeated calculations. For example, nearly 40% of tokens may be wasted during code review processes.
- Hidden Fees from Tokenizers
Tokenizers, which are crucial for processing text, have been updated by closed-source models (like Anthropic) to use more detailed segmentation methods, resulting in a 47% increase in token consumption. This is similar to how mobile operators may charge you more for data usage without your knowledge.
- Redundant Skills
Many skills in AI agent "skill libraries" are ineffective, but using them still incurs token costs. For instance, if a skill includes lengthy descriptions, the AI may waste tokens simply reading those instructions instead of performing any actual work.
- Inefficient Communication
Multiple AI agents collaborating often repeat unnecessary information and deviate from the task objectives, leading to additional token consumption, similar to inefficient meetings.
3. Limited Use Cases
AI's primary strength lies in programming, but it struggles to create value in other areas:
- Programming: Efficient Feedback Loop
In programming, code can be compiled immediately, allowing for rapid iteration and correction, resulting in high token efficiency. However, this is a niche use case.
- Other Areas: High Verification Costs
For example, contracts generated by AI need to be reviewed by humans, adding significant time costs. Even if robots perform well in simulated environments, they may fail in real-world applications (such as opening doors or lifting objects), requiring extensive verification.
- Consequence: Tokens Are Limited to Specific Use Cases
Outside of programming and highly digitalized tasks, tokens fail to generate sufficient value, leading to high costs with low returns.
4. Uneven Risks Across the AI Industry Chain
The risks are concentrated in the middle tier of the AI industry chain:
- Upstream Profitability
Chip manufacturers (e.g., NVIDIA) and data centers (e.g., TSMC) reap substantial profits, as companies rely on their computing power.
- Midstream Losses
Model manufacturers like OpenAI and Anthropic are still in the red despite growing revenue. They rely on financing to sustain operations and engage in circular funding patterns (e.g., buying computing power from Oracle, which in turn buys NVIDIA chips, with NVIDIA investing in OpenAI).
- Downstream Cost Cuts
Companies are starting to reduce their token usage. Microsoft has revoked the license for Claude Code, and Meta has removed token consumption metrics from its reporting. They are now setting budgets and controlling costs more carefully.
- Potential Economic Risks
The private credit market carries $3 trillion in risks, with loans based on future expectations. Data centers' demand for energy (e.g., electricity) may lead to price increases, affecting consumer spending.
5. Solutions: Technological Optimization and Practical Use Cases
To make token usage more economical, improvements are needed in both cost reduction and value creation:
- Technical Approaches:
- Efficient Context Management: Only retain essential information and eliminate redundancy.
- Skill Optimization: Remove unnecessary skills and load them only when needed.
- Model Selection: Use cheaper models for simple tasks and high-end models for complex ones.
- Budgeting: Set token limits for each AI agent to prevent inefficiencies.
- Commercial Strategies:
- Improve governance to better manage token usage, similar to how cloud services are managed.
- Target Specific Use Cases: Identify scenarios with clear validation criteria (e.g., image recognition or demand forecasting) where tokens can create value.
- Focus on ROI: Future AI tools should focus on delivering the most value with the least amount of tokens.
Conclusion
The phenomenon of "ineconomical token usage" is a temporary challenge in the early stages of AI development, similar to the inefficiencies of early steam engines. As technology improves and practical use cases are identified, tokens will evolve from costly tools into valuable assets. The key is to focus on the actual benefits they bring to businesses, rather than just their complexity.
(End of Analysis)