Multimodal LLM API Pricing: How to Choose Between Half-Price Rates and Subscription Plans
As multimodal models move into content review, support, document analysis, video understanding, and office automation, API cost becomes a real selection factor. This article compares pay-as-you-go, cache-hit, batch, and subscription pricing, and shows how traffic patterns, latency needs, and cache rates change the true bill.
As multimodal large models move into content moderation, customer support, document analysis, video understanding, and automated office work, API cost has become a major factor in technical selection. Many developers compare only the price per million input tokens and per million output tokens. Actual invoices are also shaped by cache hits, batch discounts, regional differences, subscription credits, off-peak capacity, and minimum spend.
Choosing the cheaper channel therefore requires more than the unit price. It depends on traffic volume, latency requirements, cache ratio, and call stability.
1. Define the model and what is billed
This article discusses a native multimodal large model with these typical characteristics:
- It accepts text, images, video, and files.
- It supports very long context, which suits long documents and long-video tasks.
- It uses a sparse-activation architecture so a large total parameter count does not translate directly into the cost of a single inference.
- It offers a standard API suitable for applications, workflows, and internal systems.
- Input tokens, cache-hit tokens, and output tokens are billed separately.
Multimodal input is not necessarily billed by file size. Images, video, and documents are usually parsed or tokenized first, then counted as input tokens. Files of the same size can produce very different token counts because of resolution, page count, frame rate, and content complexity.
Budget estimates should therefore use token usage from real request logs, not a simple conversion from uploaded file size.
2. Price structure across common channels
Public list prices for common pay-as-you-go channels are close. Typical standard rates are:
| Channel type | Input price | Cache-hit price | Output price | Main trait |
|---|---|---|---|---|
| Vendor platform, limited-time plan | 0.4 CNY / million tokens | 0.115 CNY / million tokens | 1.4 CNY / million tokens | Standard price temporarily at 50% |
| Vendor platform, standard plan | 0.8 CNY / million tokens | 0.23 CNY / million tokens | 2.8 CNY / million tokens | Regular online calls |
| Domestic multi-service cloud platform | 0.8 CNY / million tokens | 0.23 CNY / million tokens | 2.8 CNY / million tokens | Usually a single API entry point |
| Domestic model aggregation platform | 0.8 CNY / million tokens | 0.23 CNY / million tokens | 2.8 CNY / million tokens | Easier switching among models |
| Overseas API channel | about 1.04 CNY / million tokens | about 0.21 CNY / million tokens | about 3.45 CNY / million tokens | Priced in USD and affected by exchange rates |
CNY figures in the table are estimates only. Actual charges also depend on exchange rates, taxes, settlement method, and payment channel. For workloads based in mainland China that do not need cross-border deployment or overseas network access, overseas channels usually have no clear price advantage.
On pure pay-as-you-go rates, the temporary half-price plan is the cheapest. After that promotion ends, differences among most domestic platforms are small. Discount rules and billing method then matter more than the headline unit price.
3. Why cache hits change cost
Cache billing is easy to overlook in multi-turn chat and long-context applications.
An enterprise knowledge-base Q&A system may resend the same material on every request:
- The system prompt
- Company policies
- Product documentation
- Prior conversation context
- The user’s current question
If the first four parts stay unchanged across requests, the platform may treat them as cacheable. After a cache hit, the repeated portion is billed at the lower cache rate. Only new content and model output are billed at the normal rates.
Estimate one request as:
Request cost =new input tokens × input unit price+ cache-hit tokens × cache unit price+ output tokens × output unit price
Suppose each request has 800,000 input tokens, of which 600,000 hit the cache and 200,000 are new, and the model returns 50,000 output tokens. At standard prices:
Input cost = 0.2 × 0.8 = 0.16 CNYCache cost = 0.6 × 0.23 = 0.138 CNYOutput cost = 0.05 × 2.8 = 0.14 CNYRequest cost = 0.438 CNY
Ignoring the cache hit and treating the request as 1 million new input tokens overstates the cost. The reverse is also true: if the system has no stable prefix and the hit rate is low, the cache price is not a meaningful saving.
Cache lifetime and matching rules also matter. Channels can differ on retention time, prefix consistency, cache granularity, and eviction. Confirm those rules with a small test before production.
4. Online calls, batch jobs, and subscription credits
4.1 Online pay-as-you-go
Online calls fit real-time chat, support Q&A, interactive coding, and any flow that must return a result as soon as the user submits a request. Latency is relatively predictable and integration is simple, but the unit price is usually higher, and peak periods may hit concurrency or rate limits.
For unstable volume or projects still in validation, pay-as-you-go is usually safer. It does not require buying a large credit balance in advance, and it avoids expired credits or idle capacity.
4.2 Batch jobs
Batch processing fits:
- Overnight document processing
- Bulk summaries, tags, and classifications
- Large-scale image or video analysis
- Offline data cleaning
- Periodic report generation
If the batch API is billed at half the standard price, cost is clearly lower than online calls when an immediate response is not required. The tradeoff is uncontrolled latency. Jobs usually queue and should not serve user-facing real-time requests directly.
A practical split is two tiers: real-time requests on the online API, and non-urgent work in a batch queue. Grading tasks this way often cuts spend more than switching vendors.
4.3 Subscription credits and off-peak capacity
Some platforms sell a monthly token allowance, or let off-peak jobs consume credits at a lower multiplier. Off-peak calls might be counted at 0.3 times normal consumption.
That model suits teams with high volume, delay-tolerant work, and a clear peak pattern. Do not judge a subscription by the advertised discount alone. Also check:
- Whether monthly credits expire
- Whether credits work across models
- Whether off-peak capacity is actually available
- Whether peak periods add delay
- Whether overage automatically falls back to standard pricing
- Whether there is a minimum subscription term
Stable daily offline work can make a subscription plus off-peak capacity the cheapest option. Highly variable volume can leave purchased credits unused.
5. Five typical usage scenarios
Scenario 1: Individual development and feature validation
Assume 5 million input tokens and 1 million output tokens per month, with a low cache-hit rate.
At standard prices:
Input: 5 × 0.8 = 4 CNYOutput: 1 × 2.8 = 2.8 CNYTotal: 6.8 CNY
Price gaps among domestic channels are small. Prioritize API compatibility, documentation quality, call stability, and ease of debugging. Switching channels to save a fraction of a yuan usually is not worth the integration and maintenance cost.
Scenario 2: A mid-size online application
Assume 100 million input tokens and 20 million output tokens per month, mainly for support and knowledge-base Q&A.
At standard prices:
Input: 100 × 0.8 = 80 CNYOutput: 20 × 2.8 = 56 CNYTotal: 136 CNY
A half-price plan would bring the theoretical cost to about 68 CNY. The discount is material, but confirm when it ends and keep prices in configuration so the budget does not suddenly double when the promotion stops.
Scenario 3: Long-document analysis
Assume heavy processing of contracts, reports, and technical documents: 1 billion input tokens per month, 70% of which reliably hit the cache, and 50 million output tokens.
At standard prices:
New input: 300 million × 0.8 = 240 CNYCached input: 700 million × 0.23 = 161 CNYOutput: 50 million × 2.8 = 140 CNYTotal: 541 CNY
The priority here is not the lowest input unit price. It is raising the cache-hit rate. Fixed system prompts, stable document prefixes, and separation of dynamic content can cut repeated input cost substantially.
Scenario 4: Large-scale offline processing
If the workload is billions of tokens per month and results do not need to return immediately, compare batch prices and throughput first. Even when standard prices match, a 50% batch discount can save a large amount.
Also count waiting cost. If delay holds up human review, data publication, or downstream workflows, the paper discount can be offset by business loss.
Scenario 5: Overseas deployment or cross-region workloads
Overseas channels are usually priced in USD. Converted input and output costs can be higher than domestic CNY prices. The reason to choose them is often network location, cross-region deployment, settlement method, or compliance for overseas operations, not the lowest unit price.
If users are mainly in mainland China, cross-border latency and payment complexity can raise total cost further. If the application itself runs outside that region, network stability can matter more than the unit-price gap.
6. Building a cost-control system that lasts
First, record real token use. Do not count requests only. Log input tokens, cache tokens, output tokens, response latency, and failed retries.
Second, separate real-time and offline work. Put delay-tolerant requests on a batch queue so not every job uses the higher-priced real-time path.
Third, raise the cache-hit rate. Structure fixed prompts, knowledge-base prefixes, and repeated context so the same input is not sent again without reason.
Fourth, set budgets and alerts. Use daily, weekly, and monthly spend thresholds, and pause high-cost jobs when token use grows unexpectedly.
Fifth, keep prices switchable. Do not hard-code them in application logic. Manage input, cache, and output unit prices through a configuration center or environment variables.
Sixth, keep a backup channel. Platforms with similar prices can still differ under rate limits, outages, and peak load. A shared request wrapper lets you change the access channel without rewriting the business layer.
7. How to choose
If a clear limited-time 50% discount is available and call volume is stable, use the discounted plan, but confirm the end date and the price that follows.
For low-frequency development and testing, choose a pay-as-you-go channel with a stable API, clear documentation, and low integration cost. Do not migrate often for a tiny price gap.
If the workload repeats a large amount of context, focus on cache billing and cache-match rules.
If work can wait, a batch API is usually cheaper than real-time calls.
If monthly volume is large and stable, compare subscription credits and off-peak capacity, but first calculate utilization, expiry, and whether capacity is actually available.
Overall, optimizing multimodal API cost is not only a search for the lowest unit price. It is a redesign of how calls are made. The methods that usually work are cutting unused context, raising cache-hit rates, separating real-time and offline work, using batch discounts where they fit, and recalibrating the budget from monitoring. For developers and enterprise users, a cost model that is measurable, switchable, and auditable is worth more over time than a one-off price comparison.
