Recently, some media reported that Microsoft has revoked the internal license for Claude Code. Claude Code is an AI programming tool launched by Anthropic, which became one of the most popular auxiliary development software within Microsoft just six months after its internal launch. This was followed by a surge in token consumption and skyrocketing costs, but the quality of output was not satisfactory. After multiple considerations, Microsoft hit the brakes and directed employees towards its own Copilot CLI.

The phenomenon of disproportionate token consumption to actual output is also common among other platform companies. Uber exhausted its entire AI programming tool budget for 2026 in just four months; some Amazon employees were consuming tokens meaninglessly; Meta quietly removed the Tokenmaxxing leaderboard for internal employees, no longer encouraging unproductive token consumption. Everyone is embracing AI, but has yet to find the right approach; companies emphasize AI-native solutions but see no benefits, only increasingly long bills. I call this "token inefficiency."

Token inefficiency is the result of multiple factors, including poor internal control, limited returns on token usage, and the architectural design of the Agent itself. In the future, these issues may gradually ease with improved internal controls and continuous optimization of technical consumption. However, to turn token net gains positive, it is necessary to optimize token costs from the supply side and address how to generate actual value from token consumption in a wide range of industrial scenarios from the demand side.

1. Good products are not cheap

In the past two years, mainstream large models have rapidly iterated, and developing companies have adopted different product combination strategies based on their market positioning, leading to changes in API call prices. Model performance has significantly improved, but good products are not cheap; the calling prices for the same tier of products have also quietly increased, becoming an important reason for raising downstream users' token consumption costs.

Leaders' tiered strategy

Anthropic is one of the first closed-source model vendors to recognize that programming is the core scenario for token monetization. The main paying users of large models are developers and corporate technical teams, who are less sensitive to price and more focused on coding efficiency and quality. Gaining an early advantage in the programming business scenario allows for token premium extraction.

Therefore, Anthropic focuses on programming in its research and development. After establishing its programming capability advantage, it launched the Claude 3 series at the beginning of 2024, being the first in the industry to adopt a flagship - mid-range - lightweight product combination, achieving tiered pricing for models of the same generation while capturing both high-end and mass markets. The Opus series is positioned as a benchmark in the programming industry, anchoring the high-end market with prices of $15/$75; the Sonnet series provides cost-effective options for daily programming and office tasks; the Haiku series targets lightweight, quick interaction scenarios at affordable prices. This fine tiered division allows Anthropic to maximize profit extraction in every price band while protecting market share This pricing strategy allows Anthropic, as a technology leader, to have more competitive means and operate more flexibly. For example, upon realizing the rapid narrowing of performance gaps with competitors, they significantly reduced prices with the release of Opus 4.5, squeezing the market space of competitors. Additionally, with the release of the next-generation model Mythos Preview, they introduced a new ultra-high-end tier on Opus, raising the flagship product's price and reversing the previous trend of continuous price reductions for high-end products.

The subsequently released Fable 5 adopts the same underlying architecture and restricts certain features for security reasons, targeting a broader market with prices of $10/$50. This pricing is not only based on performance but also on the degree of security constraints, forming a three-dimensional pricing strategy of capability layering, risk layering, and pricing layering, reclaiming the premium market.

The effectiveness of this positioning strategy was fully validated between 2025 and 2026. Anthropic's annual recurring revenue soared from about $1 billion at the end of 2024 to approximately $45 billion by May 2026. More importantly, this strategy fully protects the market premium as a product power leader, relying on performance advantages to break free from the trap of price competition and complete the value closed loop of good products not being cheap.

Price Pull of the Followers

In contrast, OpenAI and Google chose a different diversification path in the early stages of large model commercialization compared to Anthropic. OpenAI invested heavily in multimodal projects like Sora in 2024; Google built an ecosystem strategy around Gemini covering multiple product lines such as search, cloud services, and Workspace. Although these investments expanded the technological landscape, they performed relatively unimpressively in office and programming scenarios due to resource dispersion. By the time they realized that programming was the main battlefield for monetizing model capabilities and attempted to catch up, they had already lost their first-mover advantage.

OpenAI's turnaround was very decisive. On one hand, they refocused on coding and agent capabilities, cutting down on resource-intensive projects like Sora; on the other hand, they followed Anthropic to establish their own layered product matrix, closely monitoring competitors one-on-one while deliberately widening the price gap between flagship models and lightweight models. The flagship high price maintains the reputation of the leading model, while the lightweight low price captures market share. The pricing of GPT 5.5 aligns with Opus 4.7/4.8, establishing a high-end price anchor equivalent to Claude Opus, while the secondary models GPT 5.4 mini and nano are significantly lower than the same-level Claude Haiku 4.5, using price to gain market share.

Google, as the core of the Android ecosystem, already has a complete commercial closed loop, making the relationships it needs to manage more complex and its actions more cautious. Gemini needs to serve enterprise customers of Google Cloud, productivity users of Workspace, and consumer experiences of search products simultaneously. Even if they realize the importance of programming, they cannot decisively focus all resources on programming and office tasks; they still need to pursue a multimodal and diversified route Google is also closely following Anthropic by dividing its products into the flagship Pro series and the lightweight Flash series starting from the 1.5 generation Gemini, but the product iteration speed is relatively slow, with a lower price positioning. The flagship model Gemini 1.5 Pro, released in early 2024, outputs a million tokens at a price of only $5 under short prompt conditions, which is one-third of the price of GPT-4o and one-fifteenth of Opus 3.

The price for outputting a million tokens for the Gemini 3.1 Pro, released in February 2026, increased to $12, significantly lower than the $15 for GPT 5.4 and $25 for Opus 4.6/4.7 during the same period. Moreover, Google has implemented a reverse operation by adding an ultra-lightweight product line Flash-Lite under the lightweight product line Flash, bringing the calling price down to the same level as open-source models, which is a typical price-for-volume strategy.

The long-awaited Gemini 3.5 Pro has yet to be officially released, reflecting the internal struggles Google faces in balancing performance, security, and ecological adaptation. The pricing strategy for the new generation flagship model is also highly anticipated by the market.

Figure 1: Pricing trend changes for flagship models. The pricing for the Claude series and GPT-4o/4.1/5.4 comes from the official pricing page; the pricing for the GPT-5.5 series and Gemini 3.5 Flash comes from OpenAI/Google platforms and third-party summaries; the GLM series pricing is based on the overseas Z.ai platform, with specific prices affected by exchange rate fluctuations and dual-track pricing. Illustration: Codebuddy

The secondary/lightweight and open-source/semi-open-source model markets have quietly increased prices amid a surge in demand.

Flagship models compete on performance, while secondary/lightweight models compete on price, which is the correct posture in market competition. In the face of fierce market competition, the general expectation is that the market price center will continue to decline. However, the reality is quite the opposite; the economic token market composed of secondary/lightweight - open-source/semi-open-source models has seen its price center quietly rise over the past two years, and the true elevation of the price floor in the token market has been completed in this upward movement.

On the surface, this appears to be a frenzied red ocean. Low-cost secondary/lightweight models like Sonnet, mini, and Flash are economical options for mainstream closed-source models targeting the mass market, primarily aimed at capturing market share. At the same time, open-source or semi-open-source models like DeepSeek, Qwen, and GLM have rapidly emerged, generally adopting a flagship positioning with secondary/lightweight pricing strategy, bringing continuous price pressure to the secondary/lightweight closed-source model market By the end of 2024, DeepSeek V3 will enter the market at a price of approximately $0.27/$1.10, significantly lower than similar closed-source models. The later released R1 offers enhanced inference capabilities at a price of $0.55/$2.19, directly compressing the pricing space of GPT-4.1 mini and Claude Haiku. GLM-4 Plus provides near GPT-4 level capabilities at just $0.69/$0.35, which is highly attractive to price-sensitive developer groups. The pricing seems to be the norm in this tiered market.

On the other hand, each generation of secondary/lightweight and open-source/semi-open-source models comes with an increase in the price floor. For example, Haiku 3.5, launched in October 2024, has an input/output pricing of $0.80/$4.00; a year later, Haiku 4.5's pricing rises by 20% to $1.00/$5.00.

Around the same time, the pricing of the GPT mini series nearly doubled, from $0.15/$0.60 for 4o mini to $0.40/$1.60 for 4.1 mini. The Gemini Flash series also saw a similar increase, from the ultra-low pricing of $0.10/$0.40 for 2.0 Flash to $0.30/$2.50 for 2.5 Flash, with the million token output pricing increasing more than sixfold.

Open-source/semi-open-source models like the GLM series have seen pricing for GLM-5 in overseas markets increase by approximately 67% to 100% compared to GLM-4.7. In the words of Zhipu, this significant price increase demonstrates that the technical capabilities and market competitiveness of domestic models are rapidly improving.

The fundamental reason for this phenomenon is the explosive growth in the consumption of economical tokens. Most everyday coding tasks, document processing, and automation workflows do not require the capabilities of Opus or GPT-5.5, but are handled by models like Sonnet, mini, Flash, or completed by open-source/semi-open-source models. With the proliferation of AI coding assistants, Agent workflows, and enterprise-level AI applications, the usage of these secondary/lightweight - open-source/semi-open-source models has surged, far exceeding flagship models.

On one hand, this leads to a rapid increase in the consumption of economical models, making the cash-burning game of maintaining low prices unsustainable; on the other hand, it also opens up pricing space for manufacturers, with demand continuing to grow rapidly even as prices rise. Therefore, even in the economical token market, the competitive logic has shifted from which token is cheaper to which token offers better value for money. Whether it is Claude Sonnet/Haiku, GPT mini/nano, Gemini Flash, or DeepSeek, Qwen, GLM series, there is a trend of rising pricing centers From the above analysis, it can be seen that the token market is undergoing a process of overall uplift characterized by a solidification of high-end pricing, a simultaneous rise in mid-range volume and price, and an economic segment following the trend. Anthropic has established the strongest pricing power in the industry with its leading coding capabilities, while OpenAI and Google are accelerating their catch-up but still need to exchange price for volume in the short term. Meanwhile, open-source/semi-open-source models are not only continuously raising the pricing floor but are also beginning to share in the market growth dividends.

The evolution of this pattern will profoundly impact the profit distribution and competitive landscape of the entire AI industry. In the token market, where consumption is surging and unit prices are rising, the corresponding explosion in revenue for model vendors inevitably leads to increased costs for downstream token users, which is the underlying reason for the economic inefficiency of terminal token consumption.

Figure 2: Pricing trends for secondary/lightweight and open-source/semi-open-source models. The pricing for the Claude series and GPT-4o/4.1/5.4 comes from the official pricing page; the pricing for the GPT-5.5 series and Gemini 3.5 Flash comes from OpenAI/Google platforms and third-party aggregations; the pricing for the GLM series is based on the overseas Z.ai platform, with specific prices affected by exchange rate fluctuations and dual-track pricing. Illustration: Codebuddy

II. The Invisible Consumption of Agents

While tokens becoming increasingly expensive certainly hurts the wallet, what is even more distressing is that many tokens are systematically wasted when calling agents to perform tasks. Context traps, tokenizer black boxes, skill redundancy, and communication taxes and long-range entropy increases in multi-agent collaboration all contribute to the internal technical roots of token inefficiency.

Context Traps

Model inference requires calculating the relationship between each token and other tokens, so the longer the context, the heavier the computational burden and the more tokens consumed. For the same question, if it is presented to the agent without context, it consumes very few tokens. However, if it includes historical dialogues, tool logs, code files, error messages, and multi-turn discussions, the input token consumption may increase by several orders of magnitude.

The agent architecture naturally amplifies the long-text trap. Agents will decompose questions, plan tool calls, read files, check feedback, modify plans, and call tools again, repeating this cycle, where each step may reintroduce historical records into the context. The same batch of information is read repeatedly, and the same task is billed multiple times.

Salim et al.'s analysis of the ChatDev framework found that the tokens consumed during the code review phase accounted for an average of 39.5% of total consumption, the highest among all development phases, indicating that nearly 40% of token expenditure occurs in the process of repeatedly passing existing information between agents rather than genuinely generating new content

Figure 3: Analysis of Token Consumption Proportions at Each Stage of 30 Tasks in the ChatDev Framework. Salim, et al., (2026). Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering. Proceedings of the Mining Software Repositories Conference (MSR).

Tokenizer Black Box

The tokenizer is the foundation of large model training, determining the upper limit of information density under the same parameter count, the lower limit of effective context length, and the reliability of edge cases. The more reasonable the tokenization, the more efficient and stable the model training and inference will be. The tokenizers and weights of open-source/semi-open-source models are usually public, while the tokenizers of closed-source models are "black boxes," and updates to the tokenizer often accompany changes in token density.

In April 2026, Anthropic replaced the underlying tokenizer when releasing Opus 4.7. According to disclosures in Anthropic's official documentation, the tokenizer adjustment primarily considered the actual needs of model training, adopting a finer-grained subword segmentation scheme to enhance performance, with the side effect being that the number of tokens for the same length of text increased by 1.0 to 1.35 times. Results from multiple independent testing organizations showed that the actual inflation factor was even higher.

The enterprise AI cost management platform Finout's weighted real-world testing of actual enterprise prompts showed that the average inflation rate for technical documents and English-dense code files reached 1.47 times; ClaudeCodeCamp's comprehensive testing results for seven types of real files averaged 1.325 times; developer Simon Willison found through direct API comparison that the same system prompt under the new tokenizer inflated from 5,039 tokens to 7,335 tokens, while the token inflation for high-resolution images reached as high as 3.01 times.

Earlier, when OpenAI released GPT-4o, it upgraded the tokenizer from cl100k_base to o200k_base, nearly doubling the vocabulary size, with the official explanation stating that this move aimed to improve compression rates and enhance multilingual processing capabilities. However, vocabulary inflation itself does not imply a reduction in token counts for the same text; in fact, for non-English content, changes in the granularity of the new tokenizer's segmentation may lead to an increase in token counts rather than a decrease Regarding whether finer granularity in tokenization can enhance model performance, there is currently a lack of systematic public evidence from model vendors. In the change log for Opus 4.7, Anthropic categorized the new tokenizer under Breaking Changes, merely describing the factual changes without detailing the technical motivations or performance gains.

Researchers in the community have pointed out that finer tokenization can theoretically enrich the model's vocabulary representation capabilities, particularly benefiting code understanding and structured data processing. However, whether this potential performance gain justifies the nearly 50% increase in costs remains an open question.

The iteration frequency of tokenizers is significantly lower than that of model updates, but it concerns the most fundamental billing standards for tokens, and the changes are hidden in technical details, making it nearly impossible for ordinary users to notice. Closed-source models are particularly opaque regarding tokenizers, which may contribute to the increased token inefficiency.

Meaningless Invocation of Skills

Skills are one of the key tools that make the Agent architecture more specialized. Some view skills as longer markdowns, others see them as folders filled with various references and operational instructions, and some understand skills as a super long structured prompt. In actual reasoning and Agent tasks, many skills are overly lengthy and complex, increasing token consumption.

Gao et al.'s large-scale empirical study of 55,315 public skills revealed how ineffective loading of skills wastes tokens. At the routing level, as much as 26.4% of skills have no routing description at all, resembling tool manuals without a table of contents, significantly increasing the probability of ineffective loading by the Agent.

At the content level, over 60% of skill content consists of background explanations or example texts rather than directly executable operational rules, with most tokens spent on reading instructions rather than performing tasks. More seriously, some skills densely reference files, injecting tens of thousands or even over a hundred thousand tokens in a single invocation, of which only a small proportion may be relevant to the current task.

Han et al.'s SWE-Skills-Bench benchmark further confirmed the limited utility of skills. This study tested 49 public software engineering skills on real GitHub projects, revealing that 39 skills did not improve pass rates at all, and the average utility increment across all 49 skills was a meager 1.2 percentage points, while token expenditure increased by as much as 451%.

Only 7 skills with specialized knowledge in specific coding domains provided meaningful performance improvements; moreover, 3 skills experienced performance declines due to version conflicts. This indicates that the utility of skills is highly dependent on scenario matching, and blind invocation will only increase costs.

The Rambling of Multiple Agents and the Drift of Long Tasks

Multiple Agents are currently a favored working method, allowing a user to lead a team composed of AI, with roles for coding, reviewing, testing, and fixing, where multiple Agents perform their respective duties and supervise each other, which indeed improves output quality in many cases However, machines can also hold ineffective meetings, repeatedly discussing previously covered task backgrounds, earlier conclusions, and formatted clichés. Each repetition consumes additional tokens, which Salim et al. refer to as the communication tax of multi-agent systems.

Moreover, delegating complex long-term tasks to multi-agent systems is becoming a mainstream practice in programming and office work, gradually expanding to everyday life scenarios such as dining and transportation. Long-term tasks inherently have a tendency to go off track. The context of such tasks is filled with tool outputs, error messages, drafts, and logs, which can easily cause the model's reasoning to gradually deviate from the target.

To correct this deviation, developers often need to add mechanisms such as summarization, memory, checks, and rollbacks, leading to more token consumption. Luo et al. observed in their study of TabTracer that traditional chain reasoning tends to fall into a loop when the path is too long, and adversarial injection can intentionally trigger this loop, causing the agent to repeatedly consume tokens on the wrong path without realizing it.

This additional consumption required to maintain stability is often referred to as entropy tax. The more complex the system, the freer the agents, the more supervision is needed, and the longer the tasks and larger the context, the faster the entropy tax grows. A seemingly efficient team of agents may spend more than half of their token bill on internal coordination and self-correction.

Context traps, tokenizer black boxes, meaningless skill calls, verbose language, and long task deviations—these factors combine to create an effect on token consumption that is not simply additive but multiplicative and exponential. More importantly, these technical losses have asymmetric impacts on different users.

Developers with a technical background can alleviate these issues to some extent by adjusting system prompts, trimming skill content, and setting context window management strategies. However, ordinary business users lacking a technical background neither understand the internal token flow mechanisms of agents nor can they effectively intervene in their behavior patterns. They only see the numbers on the bill continuously increasing, without knowing where the money is actually going or why so much is being spent.

In this sense, token inefficiency is not only a technical efficiency issue but also a technical equity issue. The threshold for using AI tools has shifted from whether one can write code to whether one can understand the cost dynamics of agent architecture. In reality, most users of intelligent agents do not have the relevant technical background and are placed at a structural disadvantage.

Three, seeking real needs

Compared to pricing, ineffective consumption, and other supply-side issues, the limitations on the application side are a more significant reason for token inefficiency. Although model performance has made remarkable progress in the past two years, the universality of tokens remains quite limited. Current token usage is mostly confined to scenarios with a high level of digitization, such as programming assistance, document processing, and data analysis.

Outside of these advantageous areas, the performance of large models sharply declines as the level of digitization in application scenarios decreases. In offline service industries with extremely low levels of digitization, such as dining, housekeeping, retail terminals, and on-site repairs, the tasks that tokens can independently complete are limited to the already highly digitized process management parts, making it difficult to participate in on-site operations This does not mean that AI will never enter these fields, but rather that there is a structural gap between the current pure language model paradigm and the real world. This issue has existed since the mobile internet era and is fundamentally why digital technology has failed to fundamentally change the primary and secondary industries.

The development of artificial intelligence provides new possibilities for bridging this gap, with foundational research in scientific intelligence, world models, robotic systems, and more making progress. In the past two years, the Nobel Prize in Physics and Chemistry has been awarded to AI scientists, and humanoid robots like Figure, Tesla Optimus, and Yushu have made significant advancements. However, these cutting-edge fields are still in the laboratory stage, and before achieving groundbreaking application-level breakthroughs, tokens are likely to remain trapped in highly digitized scenarios.

Programming is a universal special case

Programming is currently the application scenario where large language models perform best, but this scenario is not universally representative; a more accurate description is that it is a universal special case.

Universality means that programming outputs a universal language for agents, which can directly drive different types of agents to assist in completing a variety of tasks in well-digitized scenarios. From this perspective, it is not surprising that Claude Code, which specializes in programming from Anthropic, and OpenAI's GPT Codex have become the most popular agent products on the market.

The special case refers to the significant advantages that programming scenarios have in the model's post-training phase. First, there is definite signal feedback; the code generated by the model can be run, and compilers, interpreters, and unit tests can immediately provide precise, structured, and unambiguous judgments of correctness. Second, based on this automatic signal feedback, an efficient automatic post-training closed loop can be formed, allowing feedback to seamlessly enter the reinforcement learning loop, enabling agents to rapidly generate, report errors, and self-correct in a digital sandbox. Such autonomous training environments are rarely seen in other scenarios and are even fundamentally impossible to form.

Once programming is left behind, the efficiency of model training will significantly decrease. In the traditional business world, where the level of digitization is relatively low and automatic post-training closed loops cannot be formed—such as in management decision-making, legal negotiations, clinical medicine, and supply chain logistics—the costs of data collection and result verification will consume any token economy. Agents that cannot obtain low-cost feedback signals will be unable to achieve exponential self-evolution and will struggle to replicate their tremendous success in programming.

In February 2023, A&O Shearman was the first to reach an exclusive strategic partnership with the vertical large model company Harvey AI in the legal field, deploying the AI legal assistant developed by the latter across A&O Shearman's 43 offices worldwide. During a trial period of several months, over 3,500 lawyers at A&O Shearman submitted approximately 40,000 queries to Harvey, covering various legal workflows such as contract drafting, regulatory research, and due diligence, which indeed improved work efficiency On the other side of the coin, A&O Shearman clearly stated in its official press release that all outputs generated by Harvey AI must undergo careful review by practicing lawyers before they can be used. AI has not truly replaced the professional judgment of lawyers; it has merely added an AI preliminary review stage to the existing workflow. The time spent by senior partners reviewing contract drafts marked by AI is nearly equivalent to the time required to review the original contract from scratch.

Of course, the feedback from manual reviews is high-value data for subsequent model training, but the cost of such feedback is clearly much higher than that of programming an automated closed loop. It cannot be ruled out that in the future, when feedback data accumulates to a certain critical point, the performance of the agent in real-world scenarios will significantly improve, approaching or even surpassing the level of professionals. However, compared to programming, there is still a long way to go before this critical point is reached.

The Difficult Leap to the Physical World

The main content of legal work tasks still involves a large amount of text processing, which is a highly digitalized scenario that will certainly be highly digitized. As the components of work tasks that can be digitized and directly controlled and operated from the digital world decrease, the proportion of tasks that agents can complete will also decrease. Although most facilities in the real world are driven by software, relying solely on agents to write code to control the physical world still faces enormous obstacles.

Taking the development of humanoid robots as an example, although they have surpassed the best human performance in marathon races, humanoid robots still struggle with most tasks in the real world. Cleaning, transporting, opening doors, and navigating cluttered scenes are actions that are easy for humans but pose significant challenges for robots.

Thus, Moravec stated, "It is relatively easy for computers to perform at adult levels on intelligence tests or chess, but it is extremely difficult, if not impossible, to give them the perception and action capabilities of a one-year-old child." Nearly forty years later, the value of this statement is still increasing.

In her lengthy article "From Words to Worlds," Fei-Fei Li lists spatial intelligence and embodied intelligence as mid-term goals that will take longer to mature. The reason is that the real world has no compiler; the physical world does not accept iterations, only validations, and the cost of validation is always higher than the cost of generation.

The simulation technology that was once highly anticipated has had some effect, but there is still a long way to go to achieve performance similar to that of agent adaptation in programming scenarios. Simulation technology is designed to bypass the problem of the physical world having no compiler, creating a virtual validation space using digital twins and physical engines. However, the development of embodied intelligence still encounters the gap between the virtual and the real; the optimal control trajectories trained in simplified sandboxes with massive tokens become extremely fragile when faced with real-world friction, material fatigue, and environmental noise.

Aljalbout et al. believe that the gap from simulation to reality is not a single issue, but rather a combination of multiple sub-gaps, including dynamic differences, perceptual distortions, actuator nonlinearity, and system design flaws, making perfect simulators computationally infeasible In addition, simulation training strategies often achieve inflated performance by utilizing inaccurate but certain boundary conditions in modeling. However, when deployed in real environments, these strategies are often unreliable and can even pose risks. For example, OpenAI's Dactyl dexterous hand project accumulated training experience equivalent to 13,000 years of work in simulation using 64 NVIDIA V100 GPUs and 920 32-core CPU servers, achieving a very high success rate in manipulating blocks.

However, the dexterous hand's robustness quickly declines when faced with non-predefined materials, temperature, and wear variations in the real world. In 2021, OpenAI disbanded its entire robotics team. Co-founder Wojciech Zaremba explained this decision by stating that resources needed to be redirected to areas where achievements could be more easily attained. Although the official stance did not list the Sim-to-Real Gap as a primary reason, the industry widely believes that the contradiction between the high computational costs of simulation training and the uncertainties of real deployment is one of the significant factors prompting OpenAI to abandon the robotics direction.

Validating model performance in the real physical world incurs time and capital costs several orders of magnitude higher than in the virtual world, and such real testing cannot be replaced. This asymmetric validation cost highlights the specificity of programming scenarios; algorithms are not omnipotent, and tokens are not either.

If the effective application range of tokens remains limited to programming and a few digital scenarios for an extended period, and cannot bridge the gap from the digital world to the physical world, the sustainability of AI industrialization and industrial AI will raise significant doubts. The future of the token economy depends on whether we can expand the effective range of tokens from digital islands to the broader real world. Before the real demand in the physical world erupts, token inefficiency may persist for a long time.

IV. The Spillover Risk of Token Inefficiency

The distribution of token inefficiency across the entire AI industry chain is not balanced. Upstream infrastructure and hardware manufacturers are profiting immensely from the current fixed asset investment boom; midstream model manufacturers are still competing on product performance, with high capital expenditures squeezing cash flow; downstream application effects vary by person and scenario, with most enterprises still holding back and observing. The risk in the industry chain is accumulating in the midstream, where model manufacturers are establishing small circles of cyclical financing in the capital market. Once the continuously accumulating risk of token inefficiency erupts, it will inevitably impact the financial market and even affect the stability of people's livelihoods.

The Uneven Distribution of Industry Chain Risks

The Token-Agent craze has driven massive funds into upstream data centers, networks, chip manufacturing, and power and energy infrastructure. TSMC's capital expenditure is expected to reach $52 to $56 billion in 2026, while Microsoft, Alphabet, Amazon, and Meta's combined AI infrastructure investment from 2025 to 2026 will far exceed $300 billion and approach the $700 billion level The midstream large model manufacturers are the engines of this round of AI investment wave, the anchor points of all optimistic expectations about AI, and the "hope of the whole village." However, although the main manufacturers have seen explosive revenue growth, they are still deeply trapped in losses, with high procurement costs for computing power. OpenAI is expected to only become profitable around 2030. Meanwhile, downstream enterprise users that are truly using agents to work and burning tokens have begun to control costs. After all, without seeing reasonable returns, setting budget limits for tokens, attributing costs, and tightening usage permissions are all logical management actions.

We compared the changes in free cash flow of representative listed companies in the AI industry chain over the past two years and the net profit margins of the most recent year. In 2025, upstream companies like TSMC and NVIDIA not only have higher net profit margins but also achieved rapid growth in free cash flow of 14.5% and 58.8%, respectively.

In contrast, downstream companies like Amazon, Microsoft, and Meta, although their net profit margins have remained stable or even improved compared to previous years, saw their free cash flow decline by 76.6%, 14.8%, and 3.4%, respectively, mainly due to a significant increase in capital expenditures. The token gold mine has yet to be explored, and those mining for gold are still investing money, while those selling shovels have already made a fortune.

This situation has historically repeated itself multiple times. In the early stages of industrial revolutions, as new technologies emerged, demand first exploded on the investment side and upstream of the industry, with massive capital expenditures in the midstream turning into huge profits upstream, while downstream final consumption was still in its infancy and insufficient to support the capacity expansion of midstream companies. Risks converge in the midstream of the industry, with capital and capacity running ahead of real payment demand.

In the short term, valuation adjustments, idle capacity, and the exit of some participants are almost unavoidable; in the long term, as long as the underlying demand ultimately takes shape, the data centers, chips, and networks built in advance will still find their place and become the productive base supporting economic growth. For the general public and regulators, it is necessary to guard against the risk of the industry chain transmitting outward through financial markets, as risk spillover can lead to significant economic fluctuations.

Figure 4: Comparison of free cash flow growth rates and net profit margins of upstream and downstream in the AI industry chain (FY2025-2026) Data source: annual reports of each company, 10-k SEC filing. Charting: Codebuddy

Revolving Financing and Shadow Credit

The risks in the industry chain are concentrated among midstream model manufacturers, while some midstream model manufacturers engage in revolving financing with upstream hardware companies, making it difficult to discern whether it is real growth driven by technology or a valuation game supported by capital self-circulation For example, the "AI perpetual motion machine" formed by OpenAI, NVIDIA, and Oracle starts with OpenAI receiving strategic investment from NVIDIA, then OpenAI uses the funds raised to purchase cloud services from Oracle, and finally, Oracle uses OpenAI's payment commitments to enhance credit, issuing bonds to finance the purchase of GPUs from NVIDIA for data center construction, completing the funding loop. Each step seems to have reasonable business logic, but each step feels overly "advanced."

OpenAI's total procurement framework for computing power has exceeded $1 trillion, which is not aligned with its current annualized revenue of $33 billion, and is entirely based on expectations of high growth in the future. If downstream token terminal consumption cannot bring exponential growth in revenue for model vendors, the "commitment" will turn into a "bubble."

The expectations for token terminal consumption do not seem optimistic. According to Bain & Company, to absorb the additional 200GW of computing power by 2030, terminal consumption needs to generate approximately $2 trillion in new revenue each year. Even accounting for cost savings brought by AI, there remains a gap of about $800 billion.

Such a cycle of financing games was also seen during the internet bubble era at the turn of the century, but today's valuation bubble has half of its risks hidden in the opaque private credit market, making it harder to grasp potential risks accurately. The Federal Reserve's interest rate hikes have raised the interest rates for high-risk bond markets such as startups and leveraged buyouts, forcing banks to withdraw from this market under Basel Accord requirements, leaving space for private equity firms and ultimately giving rise to a private credit market in the U.S. of about $3 trillion.

Asset management institutions like Apollo, Ares, Blue Owl, KKR, and Blackstone provide leveraged financing for data center construction through BDCs and direct loans with terms of 20-30 years. These loans are often negotiated privately and priced using models, which may lead to term mismatches. Additionally, due to model vendors lacking cash, interest is often paid in kind, resulting in compounded risks that are not easily detectable.

A report from the Bank for International Settlements mentions that the upward potential of the AI industry chain has been fully priced in the primary and secondary equity markets, but the debt market has not yet priced in the downside risks. Once downstream demand is released slowly and revenue falls short of expectations, the valuation logic of cyclical financing will collapse, forcing a reevaluation of models in private credit, increasing the risk of bubble bursts and simultaneous declines in stocks and bonds.

Resource scarcity squeezing other demands

The expansion of computing power driven by token consumption has created an extreme thirst for resources such as water and electricity in data centers, often creating a significant supply gap in the short term, exerting pressure on the local population's access to water and electricity.

The data center alley in Northern Virginia, USA, concentrates the highest density of data centers globally, carrying about 70% of global internet traffic. Due to local power grid capacity being locked in by technology companies through long-term wholesale agreements, the energy quotas for residents and traditional businesses have been severely compressed According to a report released by the Joint Legislative Audit and Review Commission of Virginia in December 2024, the electricity consumption of data centers has exceeded more than twice the output of Virginia's largest nuclear power plant. To meet the energy demands of planned or under-construction data centers in Loudoun County, it will require adding generation capacity equivalent to several nuclear power plants to the grid by 2030.

The frenzied purchasing of high-voltage transmission lines and clean energy by data centers has forced local utility companies to invest heavily in upgrading the grid. Dominion Energy plans to invest billions of dollars in grid expansion over the next fifteen years. This massive infrastructure cost will ultimately be passed on to residents' monthly bills in the form of grid maintenance fees, capacity charges, and more.

The capacity auction prices within Dominion's service area have skyrocketed from $29/MW-day to $444/MW-day, an increase of over 1400%, directly reflecting the severe scarcity of generation and transmission capacity in the grid. An analysis by the Piedmont Environmental Council of Dominion Energy's integrated resource plan indicates that during the coverage period of this plan, the electricity bills for ordinary residents could double.

The crowding-out effect of computing power expansion on everyday demand is not limited to Virginia; major global computing nodes such as Dublin in Ireland, Jurong in Singapore, and Guizhou in China have all experienced similar contradictions. In this sense, the inefficiency of tokens is not only present in the digital world but also casts a long shadow in real life.

V. Finding the Token Value Equation

Tokens are one of the most fundamental production factors in the intelligent era. Like all other production factors such as land, data, capital, and labor, as long as there is resource misallocation and factor waste, there will inevitably be so-called "inefficiencies." In this sense, the inefficiency of tokens will not just be a temporary phenomenon at the initial outbreak of the AI industry chain but will coexist with the token economy, permeating the entire development of the intelligent economy. At present, the token economy has not fully manifested, so the inefficiency of tokens is relatively prominent.

The existence of inefficiency does not mean it should be left unchecked; efforts can be made from both the supply and demand sides to reduce token inefficiency and strengthen the token economy, allowing technological development to truly translate into tangible economic value. The supply side can reduce the unit token cost through refined technical means, plugging leaks and preventing risk diffusion; the demand side can continuously explore new application scenarios to realize the value of tokens. When the downward cost curve on the supply side intersects with the upward value curve on the demand side, the net benefit after the token economy and inefficiency offset can turn from negative to positive.

Refined Technological Transformation

Context caching and semantic compression. Context caching has become a common practice among model vendors. When multi-agent pipelines frequently hit historical caches, the billing for input tokens is significantly reduced. However, this approach also has limitations; in complex enterprise-level deployments, the dispersion failure of caches caused by highly branched Agent paths results in relatively limited actual cost savings The more fundamental solution lies in context compression, which is not simply sliding to truncate historical information, but rather actively compressing at the semantic level, retaining key instructions and reasoning links while removing duplicates and redundancies. This semantic context compression can significantly reduce input token consumption while protecting instruction adherence rates.

Skill optimization and subtraction thinking. The SkillReducer research by Gao et al. provides two paths for skill optimization. One is description compression, which supplements concise information for skills lacking routing descriptions, compressing redundant background explanations and examples; the other is progressive loading, which does not load the complete skill into the context all at once, but rather loads it as needed, achieving a 39% compression of skill body.

When both are combined, while significantly reducing the token consumption of skill calls, the model's functional quality actually improves by 2.8%. This shows that more agent skill calls are not necessarily better; the benefits of subtraction when necessary far outweigh those of addition. Reducing ineffective information in the context not only lowers token consumption but also enhances the accuracy of model outputs. Less is more here not only aligns with the beauty of code but also makes tokens more economical.

Model routing and task diversion. The use of large models for simple tasks is one of the significant reasons for token waste. Adaptive model routing based on task complexity can offload simple, high-frequency sub-tasks to open-source lightweight models with specific domain capabilities, only utilizing the expensive Frontier model at critical decision points. This layered calling can significantly lower the average token cost per task without sacrificing the quality of key links.

Hard budget constraints and host architecture for multiple agents. The probability of a multi-agent system without division of labor, budget limits, and clear stopping conditions evolving into a marathon-style tea party greatly increases. The solution is to design a host architecture with hard budget constraints and asynchronous arbitration mechanisms within the multi-agent collaborative network.

The Monte Carlo tree search method proposed by Luo et al. incorporates tool verification of intermediate steps in multi-agent processes, preserving candidate states and rolling back when necessary. This idea can be elevated from the reasoning level to the architectural level, setting token budget limits for each sub-task, with the host agent monitoring global consumption and forcibly terminating ineffective loops before the budget is exhausted. This not only prevents financial loss but often also enhances the overall efficiency of the system.

Commercial value anchoring

Token governance and cost discipline. Microsoft has restricted Claude Code, and Meta has removed the token consumption leaderboard; major companies have shifted from merely encouraging token consumption to emphasizing token output and cost discipline. Quotas, approvals, model routing, cost attribution, team billing—these measures are likely to become fundamental methods of corporate AI governance in the future.

This is an inevitable stage after AI enters production systems. Even though AI is a powerful tool for promoting innovation and accelerating production, it is essential to keep track of the accounts. How many tokens were used, how much verifiable output was generated, and how much rework was caused must all be measured Without measurement, there is no management; without limits, there is no discipline. The truly advanced companies do not assess based on using AI the most, but rather on completing the most work with the least tokens.

Rationing will become the norm. Companies will not supply tokens indefinitely, but will instead set budget pools and approval processes, similar to managing cloud computing resources. This governance does not contradict technological innovation; on the contrary, rationing will compel architects to design more efficient agent systems that internalize cost constraints.

Finding realistic scenarios for large-scale commercial applications of tokens is fundamental to achieving positive net revenue from tokens. Programming and agent architecture are just small steps toward the token economy; identifying business scenarios that can generate significant productivity leaps is a prerequisite for entering the fast lane of token economic development and creating substantial economic value. Currently, there are still few cases of large-scale application of agent architecture in real business scenarios that yield significant benefits, and most are isolated cases. General solutions that can be widely applied across other companies and industries are still in development.

Embodied intelligence and digital twins are one of the expansion directions, but we must face the asymmetric validation costs brought by the Sim-to-Real Gap. A more pragmatic path is to find intermediate areas in traditional industries that have weak certainty feedback, such as imaging screening in auxiliary diagnosis, demand forecasting in supply chains, and initial contract screening in the legal field.

Although the validation costs of these scenarios are not as close to zero as compilers, they are far lower than pure physical world validation and are expected to become a bridge for the token economy to transition from the digital sandbox to the real world. OpenAI's recent resumption of robotics research indicates that while embodied intelligence is challenging, it cannot be bypassed.

Returning to ROI

Any investment that creates value exceeding the costs incurred, regardless of how advanced the technology is, will ultimately be unsustainable. The token economy's inefficiency is not a technological failure but a temporary dilemma often encountered when technology moves toward mass production. Just like the steam engine in the early industrial revolution, which was inefficient and consumed coal excessively, this does not negate the steam engine's representation of the future direction of productivity development. By continuously improving thermal efficiency and expanding application scenarios, steam power ultimately became the fundamental force driving the first phase of the industrial revolution.

Today's tokens and agent architectures are like the early steam engines: noisy and fuel-inefficient, but they have already demonstrated potential far exceeding human capabilities in specific scenarios. Their subsequent development will inevitably involve a series of technological innovations transitioning from rough to refined. The more valuable agents in the future will not be the ones with the most complex thought chains, but rather the agents that accomplish tasks using the least tokens.

As the industry transitions from a stage of flaunting technology through quantity to a production stage valuing precision, when every token consumed must justify its output value, the token will return to the gold standard of ROI, and the era of agents will find its value equation