How to Evaluate LLM API Pricing: Unit Rates, Subscription Plans, and True Cost
A domestic LLM vendor recently raised its flagship model's API prices severalfold, sparking debate over whether the new pricing is justified. This article breaks down a practical framework for evaluating any LLM API pricing scheme, covering four dimensions: billing structure (input/output rates, cache hits, long-context tiers), subscription plan leverage, token efficiency per task, and a step-by-step cost modeling process for enterprise model selection.
A domestic Chinese LLM vendor recently released a new-generation flagship model whose API pricing rose severalfold over the previous version: roughly 100 RMB per million output tokens, 20 RMB per million input tokens, and about 1 RMB for cache hits. That price point now approaches or matches the pricing of leading international models, triggering widespread debate in the community. Supporters argue the model’s benchmark results rank among the best globally, so “the price is justified”; skeptics, after calculating real-world usage costs, contend its overall cost-performance ratio isn’t actually competitive.
Behind this debate lies a more universal technical question: when evaluating an LLM API pricing scheme, how do you scientifically determine whether it is actually “expensive”? This article analyzes the problem from four dimensions.
1. Unit Price Is Just the Surface: Start with the Full Billing Structure
Many developers’ first instinct when evaluating cost is to compare “price per million tokens.” But unit price is only the tip of the iceberg. A complete billing structure involves at least three elements:
1. The input-output price gap. The industry norm is for output pricing to be significantly higher than input pricing—typically 3 to 5 times. This is because the compute cost of autoregressive generation far exceeds that of one-time prompt encoding. The new model’s structure of 100 RMB output and 20 RMB input, a 5x gap, is standard practice and unremarkable in itself.
2. Cache hit pricing. A cache hit rate of 1 RMB means that for repeated system prompts and long shared context prefixes, hitting the cache can reduce cost by 95%. For agent-style applications—where system prompts often run into thousands of tokens and repeat constantly—cache hit rate directly determines the size of your bill. Before comparing prices, you must estimate your own workload’s cache hit rate; otherwise the comparison is meaningless.
3. Long-context and tiered pricing. Some vendors charge higher unit rates for very long context requests, or bill overage usage at elevated rates. These details are often buried deep in billing documentation, yet they can double your actual costs.
2. Subscription Plan “Leverage”: The Hidden Real Discount
A particularly insightful observation from the original discussion: don’t just look at the API list price—look at the actual usage quota a subscription plan provides.
Take the international market as an example. Major vendors’ subscription products generally offer a usage leverage of 10x or even 20x—meaning a $20/month subscription corresponds to API list-price quota worth several hundred dollars, while a $200 tier may unlock quota worth over ten thousand dollars. The logic behind such aggressive pricing: vendors use high leverage to attract heavy users and build usage habits, betting that the vast majority of subscribers will never actually exhaust their quota.
By comparison, Chinese vendors’ subscription leverage is generally conservative. A 199 RMB tier might only correspond to 1,200–2,800 RMB worth of list-price usage—a leverage of 6x to 14x—with much lower quota caps. This creates a seemingly paradoxical situation: the list price matches or even undercuts international products, yet heavy users end up spending more in practice.
The takeaway: pricing evaluation must account for user profiles. Light, occasional users can rely on list prices. Heavy users—such as always-on coding agents or bulk document processing pipelines—must calculate “cost per unit of actual output,” and plan leverage is a critical variable.
3. Task-Level Cost: The Overlooked Question of Token Efficiency
The original discussion also cited an important data point: in third-party coding benchmarks, the new model’s average token consumption per completed task was close to that of leading international models. This introduces the third evaluation dimension—token efficiency.
Even at identical unit prices, two models can consume several times more tokens to complete the same task. Influencing factors include:
- Chain-of-thought length: reasoning models generate large volumes of invisible reasoning tokens, all of which are billed;
- Task success rate: solving a task on the first try versus retrying three times yields completely different costs;
- Output verbosity: some models tend to produce lengthy explanations, inflating output tokens.
A rigorous comparison therefore looks like this:
Task-level cost = (input tokens × input price + output tokens × output price) × retry factorCost per unit of effective output = task-level cost / task success rate
Only “cost per unit of effective output,” measured on the same task set under the same evaluation protocol, is a meaningful cross-model comparison metric. Discussing unit price in isolation from tasks easily leads to misleading conclusions.
4. A Cost Modeling Framework for Enterprise Model Selection
Combining the above, we recommend the following framework for teams evaluating LLM providers:
- Profile your workload: collect typical request characteristics—average input length, output length, context repetition (cache hit rate), and daily call volume;
- Model multiple scenarios: using the same workload data, calculate monthly bills under each candidate’s list price, cache price, and subscription pricing;
- Run small-scale tests: execute hundreds of real task calls to verify task success rates and actual token consumption, rather than trusting published benchmarks;
- Watch for elasticity: evaluate the cost curve under usage spikes to avoid steep overage rates beyond plan quotas;
- Preserve portability: integrate model services through a unified API abstraction layer to avoid deep lock-in with a single vendor and retain the negotiating power to switch at any time.
For teams operating in China that need to deploy LLM capabilities in a compliance-friendly manner, it is also worth evaluating the model services and compute resources offered by domestic cloud platforms such as Baidu AI Cloud, in line with local data compliance requirements.
Conclusion
Back to the original question: the new model’s price went up severalfold—is it expensive? The answer is that a price sheet alone cannot tell you. Benchmark rankings support the “worth the price” narrative, but conservative plan leverage, token consumption efficiency at the task level, and the actual utilization of caching together determine a user’s real bill. Genuine product competition ultimately comes down to enabling heavy users to complete their tasks reliably at a sustainable cost. For technical decision-makers, building your own cost model is far more valuable than following public debates about price.
