The Pricing Logic Shift in LLMs: From Cost Per Million Tokens to Per-Task Value

When Pricing No Longer Reflects Cost: Re-anchoring Model Value

Recently, a domestic AI provider significantly raised the API pricing for its flagship large language model. The new pricing sets output tokens at 100 RMB per million, input tokens at 20 RMB per million, and cache hits at 1 RMB per million. This price structure directly aligns with the standard pricing of top-tier overseas model providers. The announcement sparked widespread debate within the developer community. The focal point of the controversy is not whether the model’s technical capabilities match its peers, but rather the underlying business logic and industry signals of such a “counter-trend price increase” in a market historically defined by aggressive price wars.

The Strategic Significance of Price Anchoring: A Precise Operation on Psychological Accounts

Over the past two years, the domestic LLM API market has experienced brutal price wars. To capture developer mindshare, various vendors have continuously driven down the price per million tokens, moving from “cabbage prices” to outright “free tiers.” In this environment, the sudden emergence of a model priced at 100 RMB for output is akin to placing an undiscounted luxury item on a shelf full of discount tags. The shock is comparable to finding a bottle of mineral water priced at five figures in a budget supermarket—the immediate reaction of most people is “on what grounds?”

However, a closer examination reveals that this pricing was not decided arbitrarily. The price precisely aligns with the standard (non-promotional) pricing of a leading overseas model provider. This indicates that the domestic vendor is proactively abandoning the “cost-performance” label, opting instead to enter the market with a “value benchmarking” posture. What is more intriguing is that the overseas benchmark model was in a limited-time promotional period at the time, meaning its actual invocation cost was significantly lower than the standard rate. In other words, during this promotional window, the invocation cost of this domestic model was actually higher than its overseas counterpart.

This is clearly not cost-based pricing, nor is it purely competitive pricing; it is a form of narrative pricing. The message the vendor intends to convey is clear: my model’s intelligence level is on par with the overseas benchmark, therefore my pricing should match it. You can offer your discounts as part of your market strategy; I will maintain my price as a valuation of my own intelligence level. This posture of “refusing to be a shadow” is unprecedented in the commercialization of domestic LLMs. It signifies that as the marginal utility of price wars diminishes, leading vendors are beginning to build brand barriers based on technical prowess rather than price advantage.

The Pricing Paradox Under Open-Source Strategy: Inscribing a Value on Intelligence Before Commoditization

Even more noteworthy is that this model adopts an open-weight strategy. Following industry convention, the weight files will quickly flow to third-party inference service providers after release. Given the model’s advertised technical benchmarks, third-party providers will likely offer inference services for the exact same weights at extremely low prices—some analyses estimate as low as 30% of the official price—in a very short timeframe.

This means the official price tag of 100 RMB might only be valid for a few days. Once low-cost third-party services go online, the official pricing loses market competitiveness, retaining only a symbolic function: to preemptively “inscribe a value metric” for the model’s intelligence level before the weights are commoditized. When developers see the same model service for 30 RMB on a third-party platform, their first reaction will likely not be “the official price is inflated,” but rather “I got a bargain, because the official price says it’s worth 100.”

This is a rare “high price + open source” combination in the history of open-weight models. Traditional open-source models usually rely on low-cost or free strategies to attract community developers. This vendor, however, does the opposite: using the official high price to anchor value perception, and then leveraging the open-source strategy to drive ecosystem diffusion. Whether this strategy succeeds depends on whether the model’s actual capabilities can sustain this price anchor. Based on currently published benchmark scores, the model does indeed possess the capability to challenge top-tier overseas models, providing a degree of technical confidence for this high-pricing strategy.

The Historic Shift in Cost Focus: In the Reasoning Era, the Output Side is the True Cost Black Hole

Setting aside marketing narratives and looking at the technical cost structure, this price increase also reveals a trend overlooked by most developers: the cost center of LLM applications is irreversibly shifting from the input side to the output side.

In traditional API pricing, there is typically a several-fold gap between input (Prompt) and output (Completion) prices. However, in the era of Reasoning Models, models perform extensive internal reasoning before generating an answer—all tokens generated during this “thinking” process are billed as output. Taking this model as an example, it currently only offers a “maximum thinking intensity” mode, meaning every invocation triggers the deepest reasoning chain. For complex tasks, the number of output tokens generated might be 5 to 10 times the number of tokens in the final answer itself.

Suppose a task requires the model to output 1,000 tokens for the final answer; internally, the model might generate 5,000 to 10,000 reasoning tokens, all billed at 100 RMB per million tokens. The actual output cost per invocation is far higher than the surface numbers suggest. This also explains why the model explicitly targets Agent scenarios—these are classic machine-to-machine, high-concurrency, long-chain token consumption environments. Standard conversational interactions from a single user simply cannot sustain this cost structure.

Evaluation Paradigm Shift: From “Price Per Million Tokens” to “Per-Task Cost”

As model pricing shifts from the input side to the output side, developers’ cost evaluation models must evolve in tandem. We are accustomed to comparing different models horizontally using “price per million tokens,” but in the era of reasoning models, this metric is severely distorted. Because different models have vastly different reasoning efficiencies, thinking depths, and output compression rates, the number of tokens consumed for the exact same task can vary by an order of magnitude across different models.

A more rational approach is to adopt a “per-task cost“ evaluation method: select a standardized set of tasks, and under identical input conditions, record the actual token consumption and final price required for each model to complete the task. For this specific model, the industry estimates its average per-task cost at around $0.94. This figure is not considered expensive within the commercial closed-source model camp—a leading overseas flagship model costs about $1.04 per task, while another top model is over $1.80—suggesting it might even offer decent cost-performance.

However, within the open-weight camp, this price appears highly abrupt. For similar tasks, another domestic open-source model costs only $0.32, while a model known for its low pricing requires a mere 4 cents. Converted to per-invocation pricing, the gap can range from 16 to 20 times. For cost-sensitive developers or startup teams, this disparity can determine the life or death of a project.

The Art of Trade-offs: High Prices Have Their Reasons, but Also Their Costs

Returning to the original question: how should we view this pricing strategy that increased the price fivefold?

From a technical perspective, this pricing is not an arbitrary demand but a courageous valuation of the model’s intelligence level. In a landscape where reasoning capability has become the core competitive metric, daring to align pricing with top international closed-source models indicates that the vendor has full confidence in its model’s reasoning depth and accuracy. For applications such as Agents, complex data analysis, and deep long-context understanding, if this model can significantly reduce manual intervention and task failure rates, the per-task cost might actually be lower than frequently retrying with cheaper models.

From a strategic perspective, this is a critical signal that domestic LLM commercialization is shifting from a “traffic mindset” to a “value mindset.” Price wars are unsustainable; when all models converge in capability, the industry needs new dimensions of differentiation. By anchoring value with high prices, driving ecosystem growth through open source, and aligning with real needs via per-task cost accounting, this combination of strategies provides a noteworthy case study for the industry, regardless of its ultimate market acceptance.

However, developers facing such high-priced models should remain clear-headed: the official list price does not equal actual cost, and model parameters do not equate to real capability. Before integrating any high-priced model, it is essential to conduct a comprehensive cost simulation using your own business scenarios—including output token distribution across varying input lengths, the impact of reasoning intensity on task success rates, and cost-performance comparisons with third-party compatible services. Especially when open weights are about to diffuse, waiting a month might allow you to acquire nearly identical reasoning capabilities at a third of the price.

LLM pricing is transitioning from “pay-per-volume” to “pay-per-value,” and true value will always depend on the problems you use it to solve.