Open the price list of any AI cloud service provider, and the most conspicuous item is always the pricing for "per million input/output Tokens." This billing model, jokingly referred to by the industry as the "water, electricity, and gas of AI," has dominated the sector for nearly three years. Developers watch the usage curves on their consoles with racing heartbeats, while mocking themselves in forums as "water carriers in the digital age"—weighing the water bill with every single character they type.
However, over the past six months, a batch of unconventional AI products has quietly shifted the unit of measurement on their price tags from "Tokens" to "number of tasks completed" or even "valid deliverables." This undercurrent is leveraging something far more profound than a price list: the very metric of software value.
01. Paying for Word Count, or Paying for Value?
The logic behind pay-per-token seems flawless at first glance. As the model processes inputs and generates outputs, every step consumes GPU computing power. Charging by Token is like paying a translator by the word—fair, transparent, and quantifiable.
Yet users quickly discovered a subtle misalignment: a chasm often exists between the model's "labor input" and the user's "sense of gain."
A cross-border e-commerce operator once ran the numbers for industry media: he used 2,000 Tokens to ask the model for a competitive analysis report. The model returned 3,000 Tokens of verbose text, filled with correct but empty phrases like "overall" and "notably," with only three to five lines of truly usable, effective information. Yet he had to pay in full for these 5,000 Tokens—because the platform only recognizes Tokens, not "usefulness."
Essentially, this billing method treats the internal computing power consumption of the model as the sole pricing basis, shifting the risk of value judgment entirely to the paying party. From an economic perspective, this is a typical "cost-plus" mindset. While it might be taken for granted in the era of industrial goods, it is inevitably a poor fit for AI, whose mission is to solve complex problems.
A more hidden cost lies in the users' decision-making psychology. When every round of conversation is converted into Token consumption in real time, a kind of "metering anxiety" quietly spreads. Product managers dare not use AI for brainstorming, because divergent thinking means massive Tokens are wasted; legal counsel dare not let the model compare contract clauses item by item, because the Token count for long texts can instantly blow up the budget.
People have started living as if guarding a water meter, wincing at the water bill every time they turn the tap. This psychological cost, in turn, distorts the real usage scenarios of AI—a tool meant for exploration and trial-and-error is degraded into a cheap typewriter only dared to be used for quick and straightforward tasks.
Research institutions have reported that under the Token billing model, over 60% of enterprise users deliberately compress the length of their prompts and cut down on contextual background, which precisely sacrifices the long-range reasoning and holistic understanding capabilities that AI excels at. In other words, the meter itself changes the driving style, and the true performance of this vehicle is never fully unleashed.
02. The Immense Weight of "Results"
It is precisely amidst this collective anxiety that the "pay-for-results" model, like a novel sprout breaking through the soil, has begun to take root in certain niche tracks.
The first to test the waters was the marketing copywriting generation field. Several startup AI writing tools no longer charge by the character, but by the "usable final draft"—users submit requirements, the tool generates three versions, the user selects one for fine-tuning and exports it, and the system only bills for the finally selected piece, even pricing it in tiers based on the subsequent click-through and conversion rates of that copy. Subsequently, code assistance tools followed a similar logic: instead of counting how many lines of code were completed, they charge by "PRs (Pull Requests) merged into the main branch and passed testing."
In this process, how many intermediate versions the AI generated and how many debugging conversations were discarded are entirely absorbed by the tool providers themselves. For the users, only a clean number appears on the bill: how many problems were solved and how many valid reports were produced this month.
This shift, on the surface, is an adjustment of pricing strategy, but at its core, it is a reversal of the risk-bearing relationship. In the pay-per-token era, users are like buying vegetables in a farmers' market: after weighing and paying, they wash, cut, and cook the vegetables at home. Whether the dish tastes good or suits their palate is a matter of culinary skill; the vendor bears no responsibility. Pay-for-results, on the other hand, is more like ordering food at a restaurant—the customer only describes the desired flavor and budget. How many ingredients the kitchen toils with and how many scraps are discarded have nothing to do with the customer. If the served dish is unsatisfactory, the restaurant must either remake it or waive the bill.
AI service providers have shifted from "selling computing power" to "selling deliverables." They must truly understand the problems users want to solve, rather than mechanically responding to every prompt. This forces model vendors to invest more effort in intent recognition, multi-turn clarification, and automatic error correction, because every invalid output becomes their own cost, rather than the user's bill.
Of course, this paradigm shift is far from a smooth journey. The biggest controversy focuses on the right to define "results." What constitutes a good result? Does the open rate of the copy count? Does the runtime of the code count? And how should the adoption rate of a consulting report be quantified?
In different scenarios, standards vary wildly, easily leading to endless disputes. Some enterprises have already tried using "manual user confirmation" as a hard threshold for results to take effect, but manual confirmation itself introduces new cognitive burdens and manipulation risks. A trickier challenge comes from the technical side: to reduce their own costs, will models tend to output shorter, more conservative answers to end the conversation as quickly as possible? If "results" are simply and crudely defined as "answer length below a certain threshold," it is no different from forcing AI into mediocrity.
These controversies indicate that pay-for-results is not a panacea. It is only applicable to scenarios with clear task boundaries and relatively objective acceptance criteria. For creative divergence and exploratory dialogues, the Token model remains a more pragmatic compromise.
03. From Counters to Value Scales
However, it is undeniable that the migration from Tokens to results is reshaping the cost structure and competitive logic of the entire AI service industry.
In the past, cloud vendors competed on who had cheaper computing power and higher concurrency, which was essentially a resource race. In the future, the breakthrough for new players may lie in who has a higher "result achievement rate" and a lower average number of "invalid rounds." Leading platforms have already begun to adopt hybrid billing—charging a nominal fee per Token for basic conversations, while quoting separately for complex tasks (such as financial report analysis and legal research) based on "deliverables." This is like a power company no longer charging solely by the kilowatt-hour, but selling packages based on "lighting duration" and "cooling effect." The latter's sense of value is obviously closer to the users' real needs.
Looking at a longer cycle, pay-per-token is an inevitable product of the early industrialization of AI—when model capabilities were uneven, computing power was a scarce resource, and billing by consumption was the fairest yardstick. But today, model capabilities are rapidly approaching the baseline of "mediocre usability." What truly sets them apart is no longer the quantity of generated text, but the depth of intent understanding and the precision of delivered solutions.
As software evolves from a "tool" into a "digital employee," users naturally will not pay for the time an employee slacks off, but only for the final results placed on the desk. This revolution in metrics may be closer to the essence of business than the arms race of model parameters.
- End -