A compelling argument is put forward in the book *Token Economy* by Bai Huitian, Yuan Xiaohui, and Si Xiao, proposing that the Token is emerging as the standard unit for measuring machine intelligence in the AI era. Humanity has measured grain with bushels, oil with barrels, electricity with kilowatt-hours, and information with bits; now, machine intelligence seemingly has its own measuring stick. However, upon closer examination, a fundamental question lingers: is the Token truly worthy of bearing such significant meaning?

The most significant potential risk of *Token Economy* is the conflation of today's "water meter" for the large model industry with the future "currency" of the intelligent economy. Strictly speaking, a unit of measurement and a currency exist on entirely different conceptual levels. Once a technical metric is regarded as the core measure of value in the future intelligent economy, the industry's attention, capital, and discourse will inevitably converge around it, potentially obscuring the true drivers of value creation.

Today's focus on Tokens largely stems from the current technical structure and business models of the large model industry. When a user inputs a sentence, the model segments it into Tokens, calculates computational costs based on input, output, and reasoning, and the platform charges accordingly. For developers, Tokens directly determine API call costs. For model companies, Tokens are also crucial because they link computational expenditure, model efficiency, and revenue. However, importance does not equate to fundamental essence. The Token is, first and foremost, a technical measurement and commercial billing mechanism.

Consider several analogies: taxis charge by the mile, yet no one considers "kilometers" to be the core asset of the taxi industry. Lawyers bill by the hour, but the legal services industry has never been called the "hour economy." Cloud computing once billed by CPU hours, storage capacity, and bandwidth, yet businesses genuinely care whether their systems run reliably, with few viewing CPU clock cycles as the fundamental unit of the digital economy. The telecommunications sector followed a similar trajectory. Long-distance calls were billed per minute in the last century, mobile data shifted to per-megabyte billing, and today, many users purchase "unlimited data plans." What operators truly sell is no longer call duration or data volume but the "always-on" connectivity and user experience.

The Token is likely to face a similar fate. It can precisely record how much work a model has done, but it may not tell us how much value the model has created. A company spending 10 million Tokens ultimately cares about what it yields: workable code, a reliable contract review, a usable design, or an autonomous agent capable of completing tasks. The same 1,000 Tokens might sometimes only compose a routine email, while at other times, they could help a programmer uncover a critical software vulnerability. One model might arrive at the correct conclusion using 5,000 Tokens, while another consumes 50,000 Tokens and still answers incorrectly. If we insist that Tokens measure intelligence, we encounter an unavoidable paradox: does consuming more Tokens indicate greater intelligence or simply lower model efficiency?

This reveals the fundamental difference between the Token and the kilowatt-hour. A kilowatt-hour is a stable physical quantity, comparable whether sourced from coal, nuclear, or solar power. Tokens, conversely, are highly dependent on model architecture, tokenizer algorithms, reasoning strategies, and task types. The same sentence can be segmented into varying numbers of Tokens across different models, and identical Token counts may correspond to vastly different capabilities in different systems. A unit of measurement that cannot consistently define "one unit" across vendors cannot serve as a cross-platform standard of value, let alone function as a true currency.

Therefore, the Token is more akin to a "water meter," recording the volume flowing through the pipe, but failing to indicate whether that water irrigated a field or was wasted down a drain. Meanwhile, the AI industry is rapidly moving beyond understanding intelligence through Tokens alone. Most users currently engage with large models by posing a question and awaiting an answer, but the rise of agents is reshaping this interaction paradigm. In the future, an increasing number of tasks will be framed as "help me plan a trip," "complete a due diligence review," "optimize my entire supply chain," or "design and execute a marketing campaign." Such tasks may involve dozens of model calls, hundreds of searches, multiple external tool invocations, switches between different models, and even collaboration among multiple agents.

For the end user, the number of Tokens consumed in this process will become as technically relevant but commercially obscure as the RPM of a car engine. What users are willing to pay for is the outcome. If an agent produces a report that would have previously required two weeks of work from ten consultants, the enterprise will not care whether it used 20 million or 50 million Tokens; it will care whether the report is trustworthy, can directly support decisions, and whether costs are genuinely lower compared to past approaches. As AI transitions from answering questions to executing tasks, pricing mechanisms will evolve as well. The market may shift from per-Token billing to per-task pricing; from per-call charges to pay-for-results; and eventually, to pricing based on cost savings, revenue generation, or probability of success.

This transition is already emerging. Some legal tech companies charge per contract reviewed rather than per API call. Certain coding agents experiment with billing per merged Pull Request rather than per Token. The unit of account is gradually climbing from underlying resources to delivered outcomes. This is a recurring pattern in maturing industries. In the early days of technology, people pay for underlying resources; as technology matures, pricing gravitates toward final value. Cloud computing initially sold CPUs and storage; now, enterprises purchase databases, office software, and complete business capabilities. Manufacturing initially involved purchasing machines; now, capacity can be acquired. AI is likely to evolve along a similar path.

Consequently, the truly important economic unit in the future may not be the Token but the task. More precisely, it will be an intelligent task defined by clear results, quality standards, and accountability boundaries. Standardizing such tasks uniformly remains challenging today. A contract review and a customer service reply require vastly different levels of complexity, and a scientific research question and a hotel booking cannot be measured on the same scale. Yet, as the AI economy matures, new task classifications, quality metrics, and outcome-based pricing systems will emerge. Who bears responsibility when a task fails? Is it the model provider, the agent developer orchestrating the workflow, or the enterprise that uses the final output? This question is nearly unsolvable today, but it will ultimately matter far more than "how many Tokens were used."

By that time, Tokens will still exist but will be relegated to the infrastructure background, much like how few people think about the bytes transmitted when streaming short videos or the engine RPM when taking a ride. Underlying measurements continue to operate but no longer determine how ordinary people understand an industry. From this perspective, *Token Economy* captures a genuine phenomenon but may depict a transitional stage as a more durable law.

The importance of the Token should not be underestimated. It has enabled AI companies to precisely convert model inference costs into business invoices and has prompted enterprises to consider machine intelligence usage efficiency. Without such underlying measurement, today's large-scale AI commercial services would be inconceivable. However, the ability to measure is merely the starting point of economic activity, not the endpoint of value. Oil is measured in barrels, but what truly transformed the world in the oil economy was not the "barrel" itself. The internet runs on bits, but its most significant business models were not ultimately built on "selling bits." Search, social media, e-commerce, advertising, and cloud services all created economic structures far beyond the underlying information units. AI will similarly not remain at the stage of selling Tokens.

Imagine a future where countless agents automatically invoke, collaborate, and settle with each other; humans may never see the Tokens. A company's procurement agent would automatically select the most suitable model, compare prices and quality across vendors, and complete the settlement. What managers ultimately see might be only two metrics: how much the task cost and how well it was executed. On that day, the million-Token prices we discuss today will quietly sink into the infrastructure layer, much like the per-megabyte data charges that are now almost never discussed.

So, the more pertinent question is: when Tokens retreat into the background, what will the intelligent economy use for pricing, and around what will it reorganize? If intelligence becomes increasingly cheap, what will the market be willing to pay a premium for? I lean towards three things: tasks, outcomes, and accountability. Tasks determine where intelligence is applied; outcomes determine how much value it creates; and accountability determines who bears the consequences when errors occur. These three elements are far closer to the essence of economic activity than the number of Tokens a model consumes. The Token has installed the first water meter for the intelligent economy, but when this system reaches full maturity, what will matter most is where this intelligence is applied and what value it creates.