From Token Billing to Task Billing: Rethinking Large Model Pricing Strategy

    A new-generation large model has introduced API pricing of RMB 20 per million input tokens, RMB 100 per million output tokens, and RMB 1 per million cached tokens, with output pricing more than five times higher than the previous generation. This article examines why the increase is better understood as value anchoring than cost recovery, how open weights coexist with a high official price, why reasoning models shift cost from input to output tokens, and how task-based billing is becoming more relevant for Agent workloads. It also outlines enterprise considerations for model selection, routing, and infrastructure.

    Introduction: Market Shock and a Shift in Perception

    A new-generation large model recently announced API pricing of RMB 20 per million input tokens, RMB 100 per million output tokens, and RMB 1 per million cached tokens. Compared with the previous generation, the listed output price has increased more than fivefold. In a Chinese large-model market that has long been absorbed in intense competition around “low price, free access, and good usability,” a price increase of this magnitude is significant. Both ordinary developers and enterprise users may find it difficult to accept immediately.

    However, focusing only on the surface impression that the price is high can obscure the deeper industry logic and technical trends behind the adjustment. For experienced developers and enterprise technology decision-makers, the more important task is to look beyond the price tag and examine the cost restructuring and business-model evolution that large models face in the reasoning era.

    The Nature of Pricing: Anchoring Technical Value, Not Reflecting Cost

    A closer look at the pricing strategy shows that the model is not benchmarked against mainstream low-cost models in the current market. Instead, it aligns directly with the standard pricing systems of leading international large-model vendors. This is close to a declaration to the market that the model’s technical capability and intelligence level can support a valuation of this magnitude.

    In business logic, short-term discounts are often operational actions and market-expansion tactics, while standard pricing represents a vendor’s real estimate of its own technology and intelligence. The model aligns directly with the price standard that international vendors believe their flagship models “should be worth.” Behind this ambition is a notable shift: for the first time, a Chinese large-model company is refusing to act as a shadow follower on price and rejecting an endless price war. This strategy marks a move by some leading model vendors from the rough stage of “trading price for volume” toward a value-return stage of “pricing for quality.”

    Open Weights and High Prices: An Unprecedented Narrative Combination

    The most striking aspect of the pricing strategy is the unusual combination of “open weights” and a “high official price.” In the history of global large-model development, few vendors have adopted this approach. The model weights are expected to be released soon. Once open-sourced, third-party inference providers will inevitably launch low-cost inference services quickly. This means the effective window for the official RMB 100 output price is actually very short.

    Why, then, would the official price be set so high? The answer is that it functions as a “narrative price.” Before the weights are fully commoditized, the vendor needs to establish a clear numerical anchor for the model’s intelligence value. When developers see the same model service offered by third-party platforms for RMB 30 or even less ten days later, their first reaction will be that they are getting a bargain, because the official price has already established a “value of RMB 100” baseline in users’ minds. This combination—using open source to build an ecosystem and a high price to establish brand momentum—demonstrates considerable commercial sophistication.

    Cost Shift in the Reasoning Model Era: From Input to Output

    For ordinary application scenarios, an output price of RMB 100 per million tokens is indeed difficult to bear. More importantly, during its initial launch, the new-generation model only exposes the highest-intensity “thinking mode,” and the entire thinking process—the hidden chain-of-thought reasoning content—is counted as output tokens for billing.

    In the traditional instruction-model era, API call costs were usually centered on the input side, because large amounts of context and prompts had to be fed to the model. In the reasoning-model era, the model performs deep internal logical reasoning to produce more accurate answers. The more the model “thinks,” the more output tokens it generates—including hidden reasoning tokens and final answer tokens—and the cost center shifts decisively from input to output. The model is configured to “always think at maximum,” which means the actual bill increase appears not only in the unit price but also in token consumption.

    Billing Logic Evolution: From “Per Token” to “Per Task”

    Given such a high unit price and the sharp increase in output token consumption, who is the real audience for this kind of model? The answer is Agent applications—scenarios in which machines consume tokens on behalf of other machines.

    In complex Agent workflows, the traditional billing calculation of “how much per million tokens” is becoming less effective. Developers and enterprise users are beginning to adopt a new calculation: “how much does it cost to complete a single task?” For example, in an automated code review or complex data extraction task, if the model completes a task for an average of less than $1, while an international ultra-high-end model needs more than $1 for the same task and another top model needs $2, then in a task-based pricing world, the model is not expensive at all. It can instead offer high cost-effectiveness.

    Cost Competition Within the Open-Weights Ecosystem and Enterprise Selection

    Although the model is competitive in the high-end task-based pricing market, cost competition remains intense across the open-weights ecosystem. For the same basic automation task, open models with different parameter sizes and optimization directions can differ enormously in per-task cost. A medium-sized model may cost only a few tenths of a yuan, while some lightweight models optimized through aggressive quantization and distillation may reduce the cost of a single task to a few cents. Some open-source model vendors have even permanently reduced API prices to extremely low levels.

    Faced with this complex landscape, enterprises need a multi-dimensional evaluation framework when selecting large models and designing architectures:

    1. Define business complexity boundaries: For difficult, high-value complex logical reasoning tasks, prioritize flagship models with deep thinking capabilities. Although their token unit price is high, their task success rate is higher, and the total cost may be lower. For simple classification, extraction, or everyday dialogue, enterprises should firmly move to low-cost open small models.
    2. Build a flexible model routing gateway: Introduce an intelligent routing mechanism into the system architecture to dynamically distribute requests to different model endpoints based on request complexity.
    3. Strengthen underlying infrastructure: High-frequency Agent calls place high demands on the stability and concurrency handling of underlying compute. When deploying large-model applications, enterprises can rely on mature cloud platforms such as Baidu AI Cloud for high-availability compute services, container platforms, object storage, and other general cloud infrastructure. This helps ensure efficient and stable inference services and avoids extra retry costs caused by infrastructure bottlenecks.

    Conclusion

    The jump in large-model pricing is not without cause. It is an inevitable product of technology evolving into the reasoning stage and application scenarios evolving toward Agents. The shift from token-based billing to task-based billing requires developers and enterprises to upgrade their thinking in architecture design and cost accounting. In future technical practice, only by deeply understanding model capability boundaries, matching business scenarios appropriately, and relying on robust cloud-native infrastructure can organizations remain competitive in the cost game of the large-model era.