LLM API Price Increases: How to Evaluate Unit Price, Subscription Value, and Cost per Task
New flagship LLM APIs often cost several times more than their predecessors, but unit price alone can lead to the wrong conclusion. This article explains why higher prices may reflect capability gains and real inference costs, then introduces a practical evaluation approach based on subscription effective-usage multipliers, per-task token consumption, and a four-step cost framework. Readers will learn how to compare models using cost per task, task success rate, and retry rate, and how to balance capability and cost through layered model usage.
1. The Phenomenon: New-Generation Model Pricing Has Risen Sharply
A notable recent industry phenomenon is that a vendor’s new-generation flagship model has API pricing several times higher than the previous generation: the output token unit price is around five times that of the previous generation, and input and cache-hit prices have also increased. This is not an isolated case. Flagship models from leading vendors generally adopt a “high price, high capability” pricing strategy, and the output prices of new-generation models are often multiples of those of the previous generation.
For developers, the first reaction is often “it has become more expensive.” But looking only at unit price can easily lead to a one-sided conclusion. Whether pricing is reasonable needs to be evaluated in the context of actual usage. Below we analyze this from several dimensions.
2. Why New-Generation Models Can Command Higher Prices
2.1 Capability Leap Creates Premium Space
If a new model ranks much higher on authoritative benchmarks, it means it has made substantive breakthroughs in complex reasoning, long-context processing, code generation, and similar tasks. For high-value scenarios such as complex code refactoring and deep analysis report generation, the output value per token is higher, so vendors have the confidence to charge a premium. In economics, this is a typical case of value-based pricing rather than cost-based pricing.
2.2 Real Increase in Inference Cost
New-generation models usually have larger parameter scales and longer reasoning chains. In particular, after introducing a deep-thinking mode, a single request may consume a large number of implicit reasoning tokens, and server-side GPU memory usage and concurrency scheduling difficulty both increase. To some extent, the price increase reflects a real increase in inference-side costs.
2.3 Competitive Pricing Benchmarking
Leading vendors often benchmark their pricing against one another: when industry flagship models generally have high output prices, followers also gain more room to set higher prices. This is also the logic behind the view that “it is not expensive, because others are more expensive.” Relative pricing and absolute pricing are two different things.
3. Unit Price Alone Misleads Decisions: Introduce the “Effective Usage” Perspective
The most valuable insight in the original analysis is this: the real measure of model cost-effectiveness is not the token unit price, but the “effective usage multiplier” of a subscription plan.
3.1 What Is the Effective Usage Multiplier?
Mainstream vendors generally use subscription plans: users pay a fixed monthly fee and receive a certain quota of model calls. The actual value of plans varies greatly across vendors:
- A $20 plan from some overseas providers, converted into pay-as-you-go equivalent usage, can reach $400–$800, a multiplier of roughly 20–40x;
- A $200 plan can even have equivalent usage reaching the tens of thousands of dollars;
- Plans in the RMB 100 range from some providers in the Chinese market usually have equivalent usage in the thousands of RMB, a clearly lower multiplier.
This means that even if Vendor A’s token unit price is three times that of Vendor B, if A’s plan multiplier is 30x while B’s is only 6x, the actual unit cost of subscribing to A may be lower.
3.2 Cost per Task Is the Core Metric
Another easily overlooked variable is per-task consumption. A stronger model may generate longer reasoning chains and more rounds of tool calls when completing the same task, causing per-task token consumption to approach or even exceed that of a higher-priced competitor. A simple conversion is:
Cost per task = average tokens per task × (input unit price × input share + output unit price × output share)
Assume a task consumes 50,000 tokens on average, with output accounting for 30%. Billing is RMB 100 per million output tokens and RMB 20 per million input tokens:
Cost = 35,000 × 20 / 1,000,000 + 15,000 × 100 / 1,000,000 ≈ 0.7 + 1.5 = RMB 2.2 per task
If the output share is higher or the reasoning chain is longer, the cost will rise quickly. Therefore, evaluation must be combined with the task profile of your own business rather than focusing only on the price table.
4. A Practical Framework for Evaluating Model Costs
Based on the analysis above, developers are advised to evaluate costs in the following four steps:
- Unit price comparison: List the input, output, and cache-hit prices of candidate models. Note that cache-hit price differences can be very significant, especially for scenarios with high-frequency repeated context.
- Multiplier calculation: Convert each subscription plan into pay-as-you-go equivalent usage and calculate the “plan multiplier.” This is a magnifying glass for the true cost of subscription users.
- Task profile measurement: Run 10–50 typical tasks on real business samples and record average token consumption, paying particular attention to reasoning tokens and tool-call rounds.
- Comprehensive decision-making: Make a comprehensive judgment based on cost per task, task success rate, and retry rate. A cheap model that requires multiple retries may ultimately cost more than an expensive model that succeeds on the first attempt.
5. Outlook for Pricing Strategy
Some argue that if providers in the Chinese market could align their plan multipliers with those of overseas leaders—for example, an RMB 100-tier plan providing thousands of RMB in equivalent usage and an RMB several-hundred-tier plan providing tens of thousands of RMB in equivalent usage—they would truly have the product strength to compete with international flagship products. This judgment points to the core contradiction in the commercialization of domestic models: their capabilities have entered the first tier, but room for commercial concessions is constrained by inference costs.
As inference optimization technologies such as KV cache compression, speculative decoding, and sparse activation in mixture-of-experts architectures continue to mature, and as domestic compute supply improves, inference costs have a long-term downward path. Only then may the combination of “high-priced flagship models + high-multiplier subscriptions” truly land. For enterprise users, a more pragmatic strategy at this stage is layered usage: use flagship high-price models for high-value complex tasks, use lightweight low-price models for daily high-frequency calls, and dynamically route traffic at the routing layer to strike a balance between quality and cost.
6. Conclusion
Model price increases are the result of both capability competition and cost structure. Looking only at a “5x increase” can lead to emotional reactions, but once plan multipliers and cost per task are introduced, the conclusion often becomes more nuanced. Developers should build a cost accounting habit based on their own business task profiles rather than being led by price tables or marketing language. After all, model selection is essentially an input-output calculation—whoever can deliver more effective output within a given budget is the true value choice.
