China's LLM Industry After the Price War: The Cost Battle Opens a New Competitive Dimension

From “Cents Per Token” to “Cents Per Task”: A Fundamental Shift in LLM Pricing Logic

Over the past two years, China’s LLM API market has experienced a massive pricing battle. In the early stage, vendors used “per million tokens” as the unit of measurement, pushing input prices from dozens of yuan down to single digits, with extreme quotes even entering a “one-yuan era.” This form of price competition, centered on token unit price, quickly lowered the barrier for enterprises to call LLMs and concentrated market attention on the simple dimension of “who has the lowest unit price.”

However, the price competition did not end with a simple “bottom bounce”—it concluded with a structural reconstruction of pricing logic. Entering the second half of 2025, although average market prices remain at historic lows, the price curve is no longer uniformly declining. Instead, it has diverged markedly by model capability tier. Budget-tier models continue to hold low prices or even drop further, while flagship models have collectively raised prices, and the price gap between the two has been actively widened. This divergence is not a return to the traditional stratification of “expensive gets more expensive, cheap gets cheaper,” but rather signals a redefinition of what “price” fundamentally means—shifting from “how much is ten million tokens worth” to “how much does it cost to complete a business task.”

The Diminishing Returns of Low-Price Strategies and the Shift in Customer Perception

Relying solely on token unit price can no longer build sustainable competitiveness. A core reason is that token unit price does not equal actual user spending. One cloud vendor unilaterally cut the API output price of its conversational model to a historic low, but users in real business scenarios often need multiple calls, context concatenation, and multi-turn interactions—ultimately, the bill amount did not decrease proportionally with the unit price drop. In other words, a low unit price is only a “victory at the list price level,” while what users truly perceive is the total cost of “completing a single business request.”

Industry monitoring data shows that even during the period of continuously declining list prices, the normalized average price based on model launch pricing had fallen to around 6.8 yuan per million tokens, yet the average user bill was far higher. This is not a statistical discrepancy but reveals the gap between API pricing and actual consumption—context length, output length, cache hit rate, and batch processing efficiency together determine the real cost users pay for each effective task.

Therefore, when some vendors began raising flagship model unit prices to the hundred-yuan level, the market did not see the expected user churn. Take a leading startup as an example: its flagship model output pricing once became the highest among domestic LLMs, yet its annual recurring revenue grew from hundreds of millions of dollars to nearly a billion dollars over the same period. The premise for users paying a higher unit price is that the model’s total consumption per task is lower—fewer calls, more precise output, and a lower failure retry rate. This precisely demonstrates that the unit of measurement for market competition has shifted from “list price” to “task cost.”

A Two-Track Landscape: Flagship Pricing Recovery and Budget Low Prices in Parallel

In the second half of 2025, the market presented a distinct two-track landscape. On one hand, budget-tier model prices continue to decline, with some vendors even offering permanent price cuts of up to 99%, clearly targeting high-frequency, low-complexity, cost-sensitive large-scale application scenarios. On the other hand, flagship-tier models have collectively raised prices, with increases ranging from several times to dozens of times, and no longer triggering industry-wide follow-the-leader price cuts. This “high-low divergence” replacing the previous “industry-wide uniform decline” means the market no longer accepts the one-directional logic that “all models should keep getting cheaper.”

Behind flagship price increases is a genuine reflection of cost structure. The training costs, inference compute consumption, context window length, and multimodal capability support of high-performance models are far higher than those of budget models. If flagship capabilities continue to be sold at low prices, vendors cannot cover R&D investment, let alone support continuous model iteration. Some vendors have introduced peak-valley pricing mechanisms, raising output prices during peak hours and offering discounts during off-peak hours—essentially refined pricing of compute resource supply and demand across different time periods, rather than simple price hikes.

Notably, flagship price increases did not force any vendor to follow suit or cut prices, in stark contrast to the previous chain reaction of “one cuts, many follow.” Market acceptance of higher-priced flagship models depends on whether the model can demonstrate significantly better completion quality and efficiency than low-priced models in specific complex tasks. When price and capability form a reasonable correspondence, price increases actually represent a restoration of market order.

Task Cost as the New Competitive Benchmark: A Measurement Revolution from “Model List Price” to “Unit Output”

In October 2025, five mainstream LLM vendors successively released new models within ten days—a dense release round widely regarded by the industry as the official end of the price war. Among the five releases, only a few vendors reduced prices for their new models; the rest kept pricing flat compared to the previous generation. But more tellingly, the competitive metrics each vendor promoted had fully shifted to “per-task cost”—the comprehensive expense required to complete a standard business task.

The change in this metric is intuitive and profound: one vendor reduced per-task cost from $0.40 to $0.26, a 35% decrease; another claimed to have cut its Agent task cost by 70%; and yet another emphasized through comparison that its per-task cost was only one-eighth of a leading international flagship model. Even for those new models that announced price cuts, third-party testing showed that their per-task cost had barely changed—because while list prices dropped, output quality and inference efficiency were adjusted simultaneously, keeping the total cost allocated to each task at the same level.

The technical logic behind this is that per-task cost depends on the combined performance of four core variables: model efficiency, inference latency, context utilization, and output quality. A model with a larger parameter count, if it can provide correct answers in fewer calls, may actually have a lower total task cost than a smaller model that requires multiple corrections. This is the essence of what distinguishes a “cost battle” from a “price war”—it does not assess the number on the price tag, but the total resources a model consumes for each unit of effective output in real business scenarios.

Technical Cost Optimization Pathways

Reducing task cost does not rely solely on increasing model parameter scale or training data—it is the result of multiple technical approaches working in concert. Currently, the industry-recognized cost optimization pathways mainly include the following:

Inference-side optimization is the first breakthrough point. Through operator fusion, quantization compression, KV Cache reuse, and dynamic batching, vendors can significantly reduce the compute overhead required for a single inference without degrading model quality. For example, sparsifying the attention mechanism in the Transformer architecture, or enabling reduced-frequency inference strategies during low-load periods, can effectively amortize the marginal cost per token.

Model architecture innovation also constitutes a key source of cost differentiation. Mixture-of-Experts (MoE) models use conditional computation to activate only a subset of expert networks during inference, greatly reducing the activated parameter ratio—so while the model’s total parameter scale is large, its actual compute is far lower than a dense model of equivalent parameter count. This architecture compresses the floating-point operations required for inference while maintaining the model’s capability ceiling, directly lowering task cost.

Intelligent scheduling and resource reuse provide cost optimization space at the system level. Mixing tasks of different priorities and latency requirements onto the same batch of compute resources, using idle periods to process non-real-time tasks, and reusing shared prefix KV caches across multiple user requests all reduce wasted computation and improve resource utilization. The peak-valley pricing mechanisms introduced by some vendors are essentially price signals guiding users to migrate non-urgent tasks to low-cost periods, achieving peak shaving and valley filling at the macro level.

How Enterprise Users Should Respond to the New Competitive Landscape

For enterprise developers currently selecting or about to select LLM APIs, the end of the price war and the start of the cost battle means the evaluation dimensions for procurement decisions need a comprehensive update. The past method of using “per-million-token unit price” as the core comparison metric is outdated—it has been replaced by comprehensive task cost calculation.

Enterprises are advised to adopt a three-tier progressive cost evaluation framework in model selection:

Tier one: evaluate the model’s per-call list price, but only as a reference starting point, not as a decision basis. Tier two: measure the actual task cost in your own business scenario—apply the model to real business data, tally the average number of calls, average token consumption, and failure retry rate for completing a specific type of task, and thereby calculate the comprehensive cost per task. Tier three: assess cost controllability and predictability, including whether cost ceiling settings are supported, whether budget alert mechanisms exist, and whether elastic scaling for sudden traffic spikes can be done without generating unpredictable cost surges.

At the same time, enterprises can adopt a multi-model mixing strategy in their technical architecture—using budget-tier low-price models for simple classification, extraction, and formatting tasks, and calling flagship-tier models for complex reasoning, long-text generation, and multi-turn dialogue. Through task grading and model matching, an optimal balance between overall cost and output quality can be achieved.

The Endgame of the Cost Battle

The transition from a price war to a cost battle is essentially an inevitable process of the market moving from extensive scale expansion into refined efficiency competition. A price war tests who can better absorb losses, while a cost battle tests whose technical system is leaner. A healthy LLM ecosystem should not have only one voice saying “cheaper and cheaper,” but should allow models at different tiers to form reasonable pricing ranges based on their capabilities and costs.

Looking ahead, task costs will continue to decline, but the driving force will no longer come primarily from vendors’ proactive price cuts. Instead, it will come from continuous improvements in model efficiency, ongoing optimization of inference systems, and further maturation of application-side engineering innovation. For developers, the upgrade in competitive dimensions means more technical dividends—the same budget can secure higher-quality, more stable model services. For vendors, the cost battle will accelerate the elimination of participants that rely solely on low-price customer acquisition without core technical accumulation, concentrating resources among teams with genuine engineering capability and innovation strength. This process will ultimately push the entire industry toward a more sustainable development trajectory.