Rising LLM API Prices: Rebalancing Cost, Value, and Ecosystem

From “Cost Per Call” to “Total Cost Per Task”

A Chinese large language model vendor recently adjusted its output pricing to 100 yuan per million tokens, with input priced at 20 yuan and cache hits at just 1 yuan — roughly five times the cost of the previous generation. Much of the community discussion has focused on the surface-level number: “How much of a price hike is this?” But what deserves closer analysis from technical teams is the structural shift behind this price signal: LLM competition is moving from a parameter-scale race toward a more nuanced contest over cost efficiency per unit of work.

To evaluate whether this adjustment is reasonable, two layers of cost perspective are needed. The first is the straightforward per-token billing rate. The second — and far more relevant for enterprise procurement — is the total token consumption required to complete an actual business task, combined with the quality of the result.

The Physics of Inference: Why High-Performance Reasoning Costs More

Industry patterns show that as models approach frontier-level capability, inference-stage compute consumption grows superlinearly. Models with stronger reasoning abilities require more forward propagation steps, longer chain-of-thought token generation, and more frequent cache verification in scenarios like long-chain reasoning, deep planning, and complex code generation. This means the underlying computational cost of a single API request can be several times higher for a new model than for its predecessor.

A commonly accepted observation in the field is that roughly half of the intelligence gain from each generation of frontier models is absorbed by increased inference compute. At the physical cost level, therefore, a 5x price increase is less a sales strategy than a “catch-up” correction to reflect actual compute costs.

Per-Task Consumption: Higher Unit Price Doesn’t Necessarily Mean Higher Total Cost

Third-party testing has compared average token consumption per task across mainstream models. The results showed that the new model’s average consumption per task was almost on par with the two most expensive overseas models in the same generation. This is a critical data point: it suggests the new model has higher “reasoning density,” with more of its thinking process internalized into the inference flow. Its per-task token consumption did not shrink proportionally to offset the 5x price increase — in fact, it nearly doubled compared to the previous generation.

So how does total cost of ownership actually change? Suppose the previous model consumed roughly 1,000 input tokens and 2,000 output tokens to complete a business task. Under the old pricing, the per-task cost would be approximately 0.07 yuan. The new model, performing the same task, might consume 1,200 input tokens and 3,500 output tokens. Under the new pricing, per-task cost rises to about 0.44 yuan — more than a 6x increase. But at the same time, task completion quality, code accuracy, and logical rigor all improve significantly.

This leads to a key judgment call: if a business scenario has a “zero tolerance” requirement for output quality — such as financial research report generation, complex SQL rewriting, or automated test case authoring — then paying a higher unit price to reduce manual review and correction cycles may actually lower overall cost. Conversely, if the workload consists primarily of short Q&A, keyword extraction, or simple classification tasks, the new model’s advanced intelligence becomes “performance overkill,” making the pricing feel steep.

The Subscription Math: The Other Side of Value Anchoring

What truly makes enterprise users hesitate is not the per-token rate itself, but the structure of subscription plan entitlements. Leading international products typically offer roughly 20–200x the paid unit price in included usage for a $20 monthly plan, while enterprise tiers (around $200 per month) provide roughly 50–100x multipliers. In other words, if you’re a Pro subscriber, the actual “economic value” you consume far exceeds the sticker price of your bill.

For the new model under discussion, based on a common 199 yuan monthly fee in China, the included reserved usage amounts to roughly 6–14x the equivalent pay-as-you-go value. Compared to leading international products, this multiplier gap is noticeable. The raw per-token rate is not unreasonable on its own, but when factoring in the “free-tier leverage coefficient” of subscription plans, the value proposition for domestic Chinese products appears relatively tight.

The issue here is not purely a matter of pricing capability. It also involves differences in operational cost structures and business model choices. Overseas vendors can afford to offer dozens of times leverage in their subscription plans because their inference infrastructure is more mature, cache hit rates are highly optimized, and a portion of subscription costs is treated as marketing expenditure to capture ecosystem share. Chinese LLM vendors, by contrast, have not yet achieved scale in consumer subscription markets, and aggressive multiplier plans could significantly widen losses.

What Should a Reasonable Price Anchor Look Like?

From a technical economics perspective, a rational LLM pricing system should satisfy three constraints. First, pricing must cover marginal inference costs and R&D amortization. Second, subscription entitlement multipliers need to lower the barrier to entry and sustain ecosystem activity. Third, the pricing anchor should map reasonably to competitors at a “similar intelligence level,” avoiding exclusion from mainstream developer stacks due to an excessively high unit-cost coefficient.

For the new model in question, if a 99 yuan monthly plan could provide roughly 2,000–5,000 yuan in pay-as-you-go equivalent value (a 20–50x leverage), and a 699 yuan monthly plan could provide approximately 50,000–100,000 yuan in equivalent value, it would generate strong pull within the developer community. Users would view it as “an intelligent productivity tool you can adopt without hesitation,” and its product capability would be highly competitive within that price range. Conversely, if subscription leverage remains below 10x for an extended period, users are more likely to perceive that “the unit price isn’t outrageous, but the value for money is far below expectations.”

Three Practical Recommendations for Technical Decision Makers

First, build a “true cost per task” evaluation framework. When benchmarking models, don’t look only at the per-million-token price. Select 10–20 high-frequency tasks from your business, run complete end-to-end tests, and track average token consumption per task, result usability rate, and manual correction time. Convert these into a “total cost to complete 1,000 business tasks” metric for apples-to-apples comparison.

Second, use model capabilities in tiers. Keep low-complexity tasks like simple intent recognition, keyword extraction, and format cleaning on lightweight models, and route only high-intelligence workloads such as complex reasoning, deep content creation, and code generation to premium models. Hybrid routing strategies typically save 40–60% of total inference cost — a more controllable approach than waiting for price cuts.

Third, don’t over-index on “latest and greatest.” The intelligence gains from frontier models are imperceptible for most enterprise workloads. For tasks like customer service Q&A and document summarization, the previous generation is often more than sufficient. Allocating budget to a “second-tier flagship” model with balanced performance and high subscription leverage often yields a better return on investment.

Conclusion: Price Is a Signal, Not a Conclusion

This round of LLM price increases is, in effect, a market-driven attempt to redefine “the reasonable price of intelligence.” Once output pricing crosses the 100-yuan threshold, the rules of the game shift from “who can give more away for free” to “who can deliver more effective intelligence per dollar spent.” For enterprise users, the metric worth watching is always the same composite formula: unit price × per-task consumption ÷ effective output quality. This latest price adjustment simply makes the denominator of that equation a little more visible.