Open the price list of any AI cloud service provider, and the most eye-catching item will always be the price tag for 'input/output per million Tokens'. This billing model, dubbed by the industry as the 'AI utilities', has now dominated the sector for nearly three years. Developers find their hearts racing as they watch the usage curves on the console, while self-deprecatingly referring to themselves as 'water carriers in the digital age' on forums—every character typed is weighed against the cost.

However, in the past six months, a group of unconventional AI products have quietly started changing the unit of measurement on their price tags from 'Tokens' to 'number of tasks completed' or even 'effective deliverables'. This undercurrent is lever (qiaodong, a Chinese term meaning 'prying open' or 'initiating change') something far more profound than a mere price list: the very metric for measuring software value itself.

01. Paying for Word Count or Value?

The logic of Token-based payment seems impeccable at first glance. The model processes inputs and generates outputs, with each step consuming GPU computational power. Charging by Token is akin to paying a translator by word count—fair, transparent, and quantifiable.

But users soon discover a subtle disconnect: there is often a gap between the model's 'labor' and the user's 'sense of gain'.

An operator of a cross-border e-commerce platform once calculated for industry media that he requested a competitive analysis report from the model using 2,000 Tokens. The model returned a verbose 3,000-Token response filled with platitudes like 'overall' and 'it is worth noting', with only three or five lines of truly usable information. Yet he had to pay in full for all 5,000 Tokens—because the platform only recognized Tokens, not 'usefulness'.

This billing method essentially treats the model's internal computational consumption as the sole basis for pricing, shifting all risk of value judgment onto the payer. From an economic perspective, this is a classic 'cost-plus' mindset, which may be justified in the industrial age but falls short when applied to AI, whose mission is to solve complex problems.

A more insidious cost lies in users' decision-making psychology. When every conversation is converted in real-time into Token consumption, a 'measurement anxiety' quietly spreads. Product managers dare not use AI for brainstorming, as divergent thinking means a flood of wasted Tokens; legal advisors hesitate to let the model compare contract clauses line by line, as the Token count for long texts can blow the budget instantly.

People begin to live as if watching a water meter, wincing at the cost every time they turn on the tap. This psychological cost, in turn, distorts AI's true usage scenarios—a tool meant for exploration and trial-and-error is reduced to a cheap typewriter used only for 'quick and easy' tasks.

Research institutions have statistics (tongji, 'statistically analyzed') that under Token-based billing, over 60% of enterprise users deliberately shorten their queries and reduce contextual background, sacrificing AI's strengths in long-term reasoning and holistic understanding. In other words, the meter itself alters driving behavior, and the vehicle's true performance is never fully unleashed.

02. The Weight of 'Outcome'

Amid this collective anxiety, the 'outcome-based payment' model has emerged like a resilient seedling, taking root in some niche segments.

The earliest adopters were in marketing copy generation. Several startup AI writing tools abandoned per-character fees and instead charged for 'usable final drafts'—users submit requirements, the tool generates three versions, and the system bills only for the version the user selects and fine-tunes for export, even tiering prices based on the copy's subsequent click-through rate. Code assistance tools soon followed a similar logic: instead of counting how many lines of code were completed, they charged for 'PRs (Pull Requests) merged into the main branch and passed testing'.

How many intermediate versions the AI generated or how many debugging dialogues were discarded were all absorbed by the tool provider. For users, only a clean number appeared on the bill: how many problems were solved or how many effective reports were produced that month.

This shift, superficially a pricing strategy adjustment, fundamentally reverses the risk-bearing relationship. In the Token-based era, users were like shoppers at a farmers' market—they paid by weight and took the produce home to wash, chop, and cook. Whether the dish tasted good or suited their palate was a matter of culinary skill, with the seller bearing no responsibility. Outcome-based payment, by contrast, resembles dining at a restaurant—customers describe their desired flavor and budget, while the kitchen's food waste and ingredient trials are irrelevant. If the dish is unsatisfactory, the restaurant must either remake it or comp the meal.

AI service providers transition from 'selling computational power' to 'selling delivery'. They must now truly understand the problems users need to solve, rather than mechanically responding to each query. This forces model vendors to invest more in intent recognition, multi-round clarification, and automatic error correction, as every ineffective output becomes their cost, not the user's bill.

Of course, this paradigm shift is far from smooth. The biggest controversy centers on defining 'outcome'. What constitutes a good outcome? Should copy open rates, code runtime, or consultation report adoption rates be quantified?

Standards vary wildly across scenarios, easily leading to endless disputes. Some enterprises have tried using 'manual user confirmation' as a hard threshold for outcome validation, but manual confirmation introduces new cognitive burdens and manipulation risks. A thornier challenge comes from the technical side: might models, to reduce their own costs, tend to (qingxiangyu, 'tend to') output shorter, more conservative responses to end conversations quickly? If 'outcome' is crudely defined as 'answer length below a threshold', it risks driving AI toward mediocrity.

These controversies suggest that outcome-based payment is not a panacea. It works best for tasks with clear boundaries and relatively objective acceptance criteria, while Token-based models remain a more pragmatic compromise for creative, exploratory dialogues.

03. From Counter to Value Scale

Undeniably, the shift from Tokens to outcomes is reshaping the cost structure and competitive logic of the entire AI service industry.

In the past, cloud providers competed on who offered cheaper computational power or higher concurrency—essentially a resource race. In the future, new entrants may break through by achieving higher 'outcome success rates' or lower average 'ineffective rounds'. Some leading platforms have already adopted hybrid billing—charging a nominal Token fee for basic dialogues while separately pricing complex tasks (e.g., financial report analysis, legal research) by 'deliverables'. This resembles power companies moving beyond per-kWh charges to bundled offerings based on 'lighting duration' or 'cooling effectiveness', which align more closely with users' true needs.

From a longer-term perspective, Token-based payment was an inevitable product of AI's early industrialization—when model capabilities varied widely and computational power was scarce, usage-based pricing was the fairest metric. But today, as model capabilities rapidly approach a 'good enough' baseline, the true differentiators are no longer text volume but the depth of intent understanding and precision of solution delivery.

As software evolves from 'tools' to 'digital employees', users naturally refuse to pay for idle time and only want to pay for results on the table. This revolution in measurement may be closer to the essence of commerce than the arms race in model parameters.

- The End -