The LLM Price War Is Over. The Real Cost War Has Just Begun

I. Aftermath of the Price War: Market Divergence Behind an Average Price Table

According to continuous tracking by industry monitoring agencies of China’s mainstream LLM API market, a highly symbolic turning point emerged in the second half of 2025. At that time, the normalized average price of domestic LLM token services once fell to a historic low—approximately 6.8 RMB per million tokens. However, this figure merely reflected the intense “price-for-volume” competition at that moment and did not reveal users’ actual billing expenditures. The market’s true direction became increasingly clear over the following two quarters: the “floor” of average prices was not continuously breached; instead, the market exhibited a pronounced “K-shaped divergence.”

On one hand, for high-frequency, lightweight invocation scenarios (such as content classification and simple extraction), prices for普惠-tier models continued to decline. Some vendors implemented permanent price cuts for specific series, with individual models seeing reductions of up to 99%, aiming to lock in long-tail traffic through absolute low pricing. On the other hand, flagship models representing the pinnacle of intelligence began to see collective price recoveries. Zhipu AI raised flagship model pricing multiple times during the year, with its CEO explicitly stating this was to “return to normal commercial value.” DeepSeek introduced a dynamic peak-valley pricing mechanism for its V4 series, significantly increasing output prices during peak hours. Moonshot AI’s K3 model set its output unit price directly at the hundred-yuan level,刷新了 the upper limit of domestic LLM pricing.

This coexistence of “affordable low prices and premium flagship pricing”宣告了 that the “price war”—which relied solely on cutting sticker prices to win customers—has ended. The market is demonstrating through action that: the commercialization of LLMs is shifting from the crude model of “selling tokens” to the refined operational model of “selling task completion efficiency.”

II. The Core of the Cost War: Redefining “Expensive” and “Cheap”

In the past, when selecting LLMs, enterprise customers often used “price per million tokens” as the most intuitive comparison anchor. But in actual business scenarios, this metric is highly misleading. A model with similar benchmark scores may require several times more token consumption when handling complex logical reasoning, long-text understanding, or multi-turn conversations. Therefore, looking only at unit price is akin to “buying the box and returning the pearl.”

The focus of current industry competition has shifted to the more pragmatic dimension of “per-task cost.” The so-called “per-task cost” refers to the total amount required to complete a specific business unit (such as generating a weekly report, handling a customer service conversation, or parsing a contract). Behind this is a comprehensive equation:

Per-task cost = Model unit price × Task token consumption + Engineering transformation and O&M costs

This shift brings three far-reaching implications:

1. Model inference efficiency becomes the core competitiveness: Vendors no longer compete solely on parameter count or benchmark scores; instead, they focus on how to reduce the generation overhead per token through architectural optimization (such as sparse attention and mixture-of-experts models), quantization compression, and efficient inference engines while maintaining equivalent intelligence levels. As one recently released model claimed, its performance surpassed many top-tier models on specific benchmarks, but more critically, the actual computational resources consumed (measured by cost) to complete equivalent tasks were significantly lower.

2. “Pareto improvement” thinking gains traction: Flagship model prices are no longer falling, but performance improvements must be sufficient to offset or even exceed the price factor. For example, when one vendor released a new-generation model, it emphasized a leap in capability on complex tasks such as software engineering. Although the unit price remained unchanged, users replacing the old version with the new one might find that tasks previously requiring 10 calls now need only 3, resulting in a lower total cost. This is the classic “price unchanged, bill halved” logic.

3. Agent costs enter the fray: As LLMs evolve from “conversation tools” to “task executors,” the unit of measurement shifts from “one conversation” to “completion of one agent task.” Pioneering players in the industry have begun focusing cost control on agent workflows. Some vendors have significantly reduced the cost of an agent executing a complex task by optimizing tool invocation and context management mechanisms. This is more commercially attractive than simply lowering model unit prices because it directly addresses enterprise users’ anxiety about “uncontrollable token consumption.”

III. Path Selection: From “Pay-per-Volume” to “Pay-per-Value”

Facing the deepening cost war, enterprise users and cloud service providers need to jointly explore new collaboration models. Currently, the industry has presented several clear evolutionary paths:

Path 1: Fine-grained Metering and Dynamic Pricing
Traditional “flat pricing” is being disrupted. In the future, we may see more tiered pricing based on task complexity, elastic pricing based on service time periods (peak/off-peak), and differentiated pricing based on service-level agreements (SLAs). This not only helps cloud vendors balance idle computing resources but also allows enterprise users to pay reasonable premiums for genuinely high-priority tasks.

Path 2: Full-Link Cost Optimization Services
For most enterprises lacking deep model tuning capabilities, simply purchasing APIs remains the mainstream approach. At this point, the value of cloud vendors lies not only in providing models but also in offering supporting engineering solutions. For example: providing more efficient vector database retrieval to reduce model dependence on context (saving tokens); providing serverless inference services so models automatically scale to zero when idle, avoiding unnecessary idle costs; and providing function computing and message queue integration so model invocations seamlessly connect with business systems, reducing application-side integration complexity and development labor costs.

Path 3: Deep Customization and Industry-Specific Models
As the price band of general-purpose flagship models stabilizes, the competitive high ground will further descend to specific industries. For high-value scenarios such as finance, healthcare, and legal services, vendors can offer industry-specific models fine-tuned on full high-quality datasets. Although such models may carry higher initial licensing or annual subscription fees, their high accuracy in specific domains can greatly reduce business losses and manual verification costs caused by erroneous reasoning, thereby winning on “total cost of ownership.”

IV. Trend Insights: Price Is No Longer the Sole Metric; Ecosystem Value Emerges

Looking back at this cycle of transition from “price war” to “cost war,” we can distill the following key trends:

• Rational return of market pricing: The high pricing of flagship models is a reasonable reflection of their substantial R&D investment and top-tier intelligence output. Enterprise users need to accept the reality that “good goods are not cheap” and shift their focus from “who is cheaper” to “who can help me make more money or save more money.”
• Urgent need for unified testing benchmarks: The current “per-task cost” still lacks an industry-recognized, standardized testing benchmark (such as comprehensive performance metrics based on TOPS—trillions of operations per second). Vendors differ in their promotional technical parameters, forcing enterprises to rely on third-party independent evaluations or self-built evaluation sets during model selection—itself a considerable cost.
• Cloud platforms become cost hubs: As model homogenization intensifies, the underlying capabilities for “running models” (computing scheduling, data storage, network latency) become hidden cost killers. A cloud platform with a cost-effective computing pool, highly optimized inference acceleration frameworks, and stable and reliable infrastructure will become a powerful backer for enterprises to win the cost war. Enterprises need to evaluate not only API fees but also the full-link cloud computing resource consumption from data preprocessing and model invocation to result storage.

V. Conclusion: Calculating a “Clear Account” for the Intelligent Computing Era

The essence of “the price war ends, the cost war begins” is the market’s shift from blindly chasing low-price bubbles to a calm examination of technical value and commercial returns. For developers, this requires enhancing “cost awareness”—considering token consumption when writing prompts, and when selecting models, looking not only at benchmark scores but also at measured business throughput and the cost-effectiveness of unit output quality.

For enterprise decision-makers, this means a completely new evaluation system: incorporating model capability, inference cost, O&M complexity, and vendor stability into a unified scorecard. Future competition is no longer a simple contest of “who is cheaper” but a comprehensive long-distance race about efficiency, quality, and long-term value. Whoever can continuously reduce the “unit cost” of intelligence through architectural innovation and management optimization in this cost war will gain the upper hand in the approaching intelligent computing era.