China’s Homegrown LLMs Enter a Paid Era: Price Tiering and Value Competition

    China’s LLM market is no longer defined only by falling token prices. Low-cost base models continue to support standardized, high-volume applications, while flagship and reasoning models are establishing higher price bands. This article examines the drivers behind the shift, explains why enterprises should evaluate total task cost rather than token unit price alone, and outlines engineering practices such as model routing, context compression, caching, output control, and validation that can help balance quality, stability, and cost. It also discusses the market’s move from price competition toward value competition based on task completion, automation, and business outcomes.

    For some time, the immediate impression of China’s homegrown LLM market has been one of rapidly improving model capabilities and continuously falling API prices. Developers could use models at very low cost for text generation, summarization, translation, customer service Q&A, and document processing. As more vendors joined price competition, costs calculated per million tokens were once pushed down to a few RMB or even lower.

    Recent public price changes, however, show that this trend is beginning to diverge. Base models remain relatively inexpensive, while flagship models have started to raise prices noticeably. The output prices of some complex reasoning models are now dozens of times higher than those of ordinary models. The conclusion is not simply that homegrown LLMs have entered a phase of “across-the-board price increases.” Rather, a clearer commercial tiering is taking shape: low-cost models support scale and broad accessibility, while flagship models serve high-value, reasoning-intensive, and complex tasks.

    1. Low-cost models remain the infrastructure for mainstream applications

    If the application scenario mainly involves summarization, rewriting, translation, information extraction, classification, customer service replies, and ordinary content generation, enterprises usually do not need the most powerful flagship model.

    At typical price levels, some mainstream models charge less than RMB 1 per million input tokens and within a few RMB per million output tokens. Assuming 1 million input tokens and 200,000 output tokens per month, the monthly model cost for a single application instance is usually only around RMB 1 to RMB 3. Even after accounting for usage fluctuations, retries, and a certain proportion of invalid requests, model cost still accounts for a relatively low share of overall system cost.

    This means that when many enterprises deploy intelligent applications, the real concern is no longer the token unit price alone. The following factors matter more:

    1. Integration and transformation costs
      Model invocation is only one part of the system. Enterprises also need to handle identity authentication, permission control, business system integration, logging, exception retries, and result write-back.

    2. Data security and compliance requirements
      When customer data, contracts, financial data, and internal knowledge bases are involved, enterprises need to clarify whether data leaves their domain, how retention periods are set, how access permissions are controlled, and whether invocation logs contain sensitive information.

    3. Output quality and stability
      Low-cost models may be sufficient for standardized tasks, but long-text understanding, complex table analysis, cross-document reasoning, and multi-turn tasks may require additional validation, retries, or human review.

    4. Overall system availability
      A low per-call price does not necessarily mean low system operating cost. In high-concurrency scenarios, rate limiting, queues, caching, monitoring, and disaster recovery all affect the final investment.

    Therefore, the core value of low-cost models is not “cheapness” itself. It is that they provide the economic foundation for sustainable deployment of large-scale, standardized intelligence.

    2. Flagship models are establishing new price bands

    Compared with base models, flagship models show a clear price gap. A typical flagship model may charge several RMB to tens of RMB per million input tokens, while output prices may reach tens of RMB per million tokens or even higher. The output price of some complex reasoning models may be dozens of times that of ordinary mainstream models.

    This difference does not necessarily mean that flagship models carry a simple “premium.” It reflects differences in inference resources and task complexity.

    First, complex tasks usually require a longer thinking process. Ordinary Q&A may only need to generate a few hundred tokens, while code analysis, complex planning, mathematical derivation, and multi-step decision-making may require much longer intermediate outputs. The more output tokens are generated, the more computing resources are consumed.

    Second, flagship models often require higher computing configurations and longer inference time. The model must not only understand the input but also perform task decomposition, result validation, error correction, and multi-round planning. These capabilities usually cannot be priced according to the resource cost of ordinary text generation.

    Third, flagship models mainly serve high-value tasks. Examples include automatically writing and debugging code, analyzing complex business materials, executing multi-step processes, and building agent task chains. In these scenarios, enterprises care less about whether a single call costs a few cents and more about whether the model can reduce manual effort, shorten processing time, and improve task completion rates.

    Therefore, flagship models should be evaluated not simply by “how much per million tokens” but by “how much it costs to complete one effective task.”

    3. Three reasons behind rising model prices

    1. The price-war phase is gradually ending

    In early market competition, low prices are often needed to quickly attract developers and enterprises. Low prices help expand invocation volume, accumulate real-world feedback, and promote the formation of an application ecosystem.

    However, once model capabilities, invocation volume, and user base reach a certain scale, continuous price cuts can compress service providers’ revenue space and may also affect investment in computing resources. To maintain R&D, inference clusters, data processing, and service assurance, vendors need to readjust their price structures.

    This does not mean that low-cost models will disappear. It means that low-price strategies will be more concentrated on base models and standardized scenarios.

    2. Inference costs are rising significantly

    The cost of large models includes not only training cost but also long-term online inference cost. Especially in complex reasoning, long-context, and agent tasks, a single request may involve multiple rounds of model calls, tool calls, state storage, and result validation.

    In a real system, one user action may trigger the following flow:

    1. User request
    2. ↓
    3. Intent recognition
    4. ↓
    5. Retrieve relevant materials
    6. ↓
    7. Generate an initial plan
    8. ↓
    9. Call business tools
    10. ↓
    11. Validate execution results
    12. ↓
    13. Generate the final response

    If the model participates in every step, the actual token consumption may be several times or even dozens of times the user input. Therefore, what appears to be a single Q&A may already have generated multiple inference requests in the background.

    3. Model competition is shifting from parameter scale to task outcomes

    As basic capabilities converge, market competition gradually shifts toward complex task completion capabilities, including code correctness, long-document understanding, tool-call success rates, process execution stability, and result explainability.

    These capabilities are more expensive to develop and operate, and they are harder to cover over the long term through simple low-price strategies. Model services will therefore form clearer tiers: base models emphasize throughput and cost, while flagship models emphasize quality, accuracy, and complex task completion rates.

    4. Enterprises should not choose models based on token price alone

    For enterprise users, model selection should be based on a comprehensive cost model rather than a comparison of input and output unit prices alone.

    The real cost of a task can be broken down as follows:

    1. Total task cost =
    2. model invocation cost
    3. + retrieval and storage cost
    4. + tool invocation cost
    5. + system operations cost
    6. + human review cost
    7. + error rework cost

    For example, a low-cost model may have very low invocation fees, but if it frequently produces format errors, factual errors, or process execution failures, the enterprise will need to add human review and retry mechanisms. The final total cost may be higher than using a stronger model to complete the task in one pass.

    Conversely, although a flagship model has a higher unit price, if it can significantly reduce rework and improve automation rates, the overall investment may actually be lower.

    Therefore, enterprises should tier tasks by type:

    • Standardized tasks: prioritize low-cost models, such as summarization, classification, keyword extraction, and fixed-format generation.
    • High-accuracy tasks: choose models with more stable quality and add structured output and result validation.
    • Complex reasoning tasks: use flagship models, with a focus on task completion rates and human-replacement effects.
    • High-concurrency tasks: focus on throughput, rate-limiting strategies, caching capabilities, and service stability.
    • Sensitive-data tasks: prioritize evaluation of deployment methods, permission systems, audit capabilities, and compliance boundaries.

    5. Engineering methods for reducing model costs

    Even if flagship models are more expensive, enterprises can still reduce overall spending through engineering optimization.

    1. Model routing

    Automatically select models based on task difficulty. Simple questions go to base models, while complex questions are escalated to flagship models. This “tiered invocation” approach is usually more reasonable than sending all requests uniformly to high-priced models.

    2. Prompt and context compression

    Long conversation histories, repeated system instructions, and irrelevant documents all increase input tokens. Summarization, deduplication, field filtering, and context trimming can reduce invalid input.

    3. Cache repeated requests

    For identical or highly similar questions, historical results can be saved. This is especially useful in knowledge Q&A, fixed-template generation, and batch data processing, where caching can effectively reduce repeated invocations.

    4. Control output length

    Output tokens are usually more expensive than input tokens. Generation length should be controlled by specifying formats, limiting word counts, and requiring bullet-point output, so that the model avoids producing unnecessary explanations.

    5. Add result validation

    Structured output, field validation, rule validation, and program tests can reduce the probability that erroneous results enter downstream processes. For code generation tasks, results should also be verified through automated tests rather than directly executing model output.

    6. The market will move from price competition to value competition

    The homegrown LLM market may show three trends in the future.

    First, base model prices will remain relatively low. As invocation scale expands and inference efficiency improves, standardized capabilities will increasingly resemble general infrastructure.

    Second, flagship model prices will not simply return to the lowest level. As long as they maintain clear advantages in complex reasoning, coding, agents, and professional analysis, they will form a relatively independent high-value market.

    Third, enterprise procurement will place more emphasis on cost per task. Market comparisons will shift from “price per million tokens” to “how much it costs to complete a business process,” “what the success rate of one task is,” and “how many manual operations the model can replace.”

    For Baidu AI Cloud and other cloud service systems, competition in model services will also expand from single-model capabilities to complete capabilities such as computing power, data management, application development, monitoring, security, and enterprise integration. What enterprises ultimately need is not an isolated chat model but an intelligent infrastructure that can connect to existing systems, run stably, and produce business results.

    Conclusion

    The entry of China’s homegrown LLMs into a paid era does not mean the end of the accessibility phase, nor does it mean that all models will become more expensive. A more accurate judgment is that the market is moving from single-dimensional price competition toward clear tiering.

    Low-cost models will continue to handle large-scale, standardized daily tasks; higher-priced flagship models will serve complex reasoning, code generation, agents, and professional decision-making. For developers and enterprises, the most important thing is not to blindly pursue the lowest unit price but to establish mechanisms for task tiering, model routing, cost accounting, and effectiveness evaluation.

    When models can truly help enterprises complete work, reduce rework, and improve production efficiency, token price will no longer be the only standard. The core competitiveness of the future LLM market will gradually shift from “who is cheaper” to “who can complete the work more reliably at a lower total cost.”