Why LLM Prices Can Differ by Several Times: A Multi-Channel Cost Guide
Quoted LLM prices for similar models can differ by several times across channels, and the highest quote is not always official direct access while the lowest quote is not always the lowest total cost. This guide applies one estimation method to a public price snapshot for four model categories, explains why gaps appear, and outlines a practical way to compare versions, billing rules, reliability, and cost per successful result.
When teams connect a large language model, many start by asking how much one million tokens cost. Once production traffic arrives, the same class of model can be quoted at several times the price through different channels. The highest price is not necessarily official direct access, and the lowest price does not necessarily mean the lowest overall cost.
Those gaps do not automatically mean a channel is pricing abnormally. Model prices are often affected by region, billing currency, context length, the input-to-output mix, caching policy, service tier, concurrency limits, and the channel’s operating costs. Comparing only one input unit price can lead to a one-sided conclusion.
This article uses a set of public pricing information and one consistent estimation method to compare the cost of four common model categories across several access channels. The figures are a public price snapshot at a specific point in time and should not be treated as long-term quotes. Before purchasing, check each channel’s current pricing page, service agreement, and billing rules.
1. Use one calculation method
To approximate real usage in AI coding, code review, automated fixes, and multi-turn agent tasks, this article uses the following assumptions:
- Input tokens account for 30% of total consumption.
- Output tokens account for 70% of total consumption.
- All figures are calculated per million tokens.
- Foreign-currency prices are converted at about 7.2 RMB per US dollar.
- Free allowances, promotions, and one-time credits are excluded from long-term cost.
- If a channel does not offer the same model version, another version is not used as a substitute.
The weighted cost is:
Blended cost = input price × 30% + output price × 70%
Output is weighted more heavily because coding tasks usually require the model to generate code, explain its process, suggest changes, and return structured results. Compared with ordinary Q&A, these tasks often consume more output tokens. If an application is mainly used for classification, extraction, or short rewrites, the input-output ratio should be recalculated.
2. Public price snapshot
The table below abstracts service sources as “official direct,” “Cloud platform A,” “Cloud platform B,” “aggregator,” and “low-cost access channel,” so the channel name itself is not treated as the conclusion.
| Model category | Official direct | Cloud platform A | Cloud platform B | Aggregator | Low-cost access |
|---|---|---|---|---|---|
| Model A | 25.20 RMB per million tokens | 22.00 RMB | 22.00 RMB | Same version not offered | 3.87 RMB |
| Model B | No unified public price | 28.80 RMB | Same version not offered | Same version not offered | 9.08 RMB |
| Model C | 5.32 RMB | 20.40 RMB | 20.40 RMB | 5.32 RMB | 1.49 RMB |
| Model D | 76.00 RMB | 76.00 RMB | No public price found | 76.80 RMB | 24.27 RMB |
On the surface, the low-cost access channel is the cheapest for all four categories, with discounts ranging from roughly 30% of the official price down to less than 20%. That only shows a difference in published list prices. It does not prove that the low-cost channel is equally strong in latency, stability, concurrency, data handling, or support response.
Model C deserves closer attention. Its blended price is about 5.32 RMB through official direct access and the aggregator, while both cloud platforms quote 20.40 RMB, close to four times the official price. “An intermediary is always cheaper than the vendor” and “a cloud platform always has a price advantage” are not reliable rules.
3. Why the same model can have a large price gap
3.1 Channels carry different costs
Official direct access usually covers the cost of operating the model service itself. A cloud platform or aggregator may also provide accounts, unified billing, access control, log auditing, network access, regional deployment, and enterprise support.
Those capabilities may matter little to an individual developer, but they can matter a great deal to an enterprise project. A higher price is not necessarily a pure markup on the model. It may include platform access, compliance management, and operational support.
3.2 Pricing units and billing rules may differ
Some services price input and output separately. Others use different prices for cache hits, long context, or batch jobs. If one quote is a base input price and another is a weighted blended price, a direct comparison will be misleading.
Before purchasing, confirm at least the following:
- Whether input and output are billed separately.
- Whether cached tokens have a separate price.
- Whether very long context carries an extra charge.
- Whether requests incur an additional per-request fee.
- Whether failed retries are billed.
- Whether streaming and non-streaming output use the same billing method.
- Whether the price includes tax or network-related charges.
3.3 Version mismatches create false discounts
Some channels may offer only a lightweight version, a coding-specialized version, or a previous generation, rather than the current flagship. Comparing those versions directly with a flagship model turns a version difference into an apparent channel-price difference.
A rigorous comparison needs three conditions:
- The model version is exactly the same.
- Context length and capability tier are broadly the same.
- Input, output, and cache billing rules are the same.
If those conditions are not met, mark the cell as “same version not offered” instead of filling the table with a similarly named model.
4. Cost patterns across the four categories
Model A: mid-range pricing with a clear channel gap
Model A’s official blended price is about 25.20 RMB, the two cloud platforms are about 22 RMB, and the low-cost channel is about 3.87 RMB. Standard channels are relatively close, while the low-cost channel creates a sharp price break.
For personal development, low-frequency calls, or prototype validation, a low-cost channel can significantly reduce the cost of experimentation. If the application involves sensitive code, customer data, or continuously running automation, price alone is not enough. Data retention, failure recovery, and service availability also need to be checked.
Model B: availability of the flagship matters more than price
Model B is a relatively new flagship tier. Because it was released recently, some channels have not listed it yet, or they offer only a nearby version with a different positioning. The useful comparison is not “who is cheaper,” but “who actually provides the same version.”
In this situation, teams should keep a version checklist that records the model name, release time, context limit, input and output prices, service region, and listing date. Price comparisons are meaningful only when those fields match.
Model C: official pricing is not always the highest
Model C’s official blended price is 5.32 RMB. Both cloud platforms quote 20.40 RMB, the aggregator is about 5.32 RMB, and the low-cost channel is about 1.49 RMB.
This shows that official direct access can be much cheaper than some resold channels. Possible reasons include different channel positioning, different platform add-on services, regional pricing, or different commercial strategies behind the published quotes.
Developers should not default to “official is expensive and a relay is cheap.” Convert prices to one unit first, then evaluate stability, quotas, and service capability separately.
Model D: higher-capability models have wider price elasticity
Model D’s blended price is close to 76 RMB, while the low-cost channel is about 24.27 RMB, a substantial gap. Higher-capability models tend to produce longer output, and complex tasks consume more tokens, so pricing errors scale quickly.
If Model D is used for a small number of high-value tasks, such as complex architecture design, critical code review, or long-document analysis, unit price may not be the main issue. If it is used for large-scale autocomplete, batch generation, or background processing, strict budget thresholds and call-routing rules are needed.
5. Unit price is not the full cost
Project cost has at least five parts:
- Token cost: input, output, cache, and consumption caused by growing context.
- Retry cost: timeouts, rate limits, and network errors can trigger repeated requests.
- Engineering and operations cost: key management, monitoring, logs, alerts, and failover.
- Integration cost: channels may differ in protocols, response formats, and authentication.
- Business risk cost: data compliance, service interruption, result-quality variation, and supplier changes.
A more useful metric is:
Cost per effective result= total call cost ÷ number of successfully completed useful tasks
A channel priced at half the others may still cost more if its failure rate and retry count are higher. For production systems, latency percentiles, rate-limit frequency, availability-zone coverage, and recovery time should also be part of the evaluation.
6. A practical selection process
Step 1: Build a version inventory
Record the exact version, capability tier, and release time of every candidate. Do not compare vague names.
Step 2: Normalize the price unit
Convert every price to RMB per million tokens, and record input, output, and cache prices separately. Do not compare different currencies or billing periods directly.
Step 3: Recalculate with real traffic
Measure the actual input-output ratio. Code generation may have a higher output share, while knowledge-base retrieval may have a higher input share. Do not keep using a fixed 30/70 split indefinitely.
Step 4: Run a small load test
At a minimum, measure:
- Time to first token.
- Full response time.
- Success rate under concurrency.
- Rate-limit and timeout rates.
- Stability on long-context tasks.
- Compliance rate of structured output.
- First-pass success rate on coding tasks.
Step 5: Set tiered routing
A practical pattern is to send simple tasks to a lower-cost model and complex tasks to a higher-capability model. Format conversion, summarization, and simple completion can use the lower-cost model. Architecture analysis, complex debugging, and final review can use the higher-capability model.
Step 6: Review prices and versions monthly
Prices, context limits, and available versions can change. Archive the pricing page, capture date, billing rules, and internal calculation sheet, and review them once a month.
7. Conclusion
Value for money in large language models is not simply “the lowest list price.” A sound choice should meet all of the following:
- The version is explicit and the comparison method is consistent.
- Pricing is transparent, with clear input and output rules.
- The service is stable under real load.
- Concurrency and quotas are sufficient.
- Data handling fits the business requirements.
- Traffic can be switched quickly if a failure occurs.
- The blended cost can be sustained within the budget.
In this snapshot, official direct access is not necessarily expensive, cloud platforms are not necessarily cheap, and aggregators or low-cost channels cannot be judged by the discount alone. For individual developers, low-cost options fit prototypes and light workloads. For enterprise production systems, stability, compliance, portability, and cost per effective result matter as much as price.
The safer approach is not to bet on a single channel. Keep one calculation sheet, test with real traffic, and design the calling architecture so it can be switched. If prices, versions, or service policies change, the setup can be reassessed quickly instead of remaining tied to one supplier.
