Multimodal LLM API Pricing Comparison and Pay-as-You-Go Cost Optimization
This guide explains why comparing only input token prices is misleading for multimodal models, summarizes typical published rates including cache hits and output tokens, and shows how discounts, caching, batch APIs, and reserved capacity change real bills. It covers five usage scenarios, a monthly cost formula, and non-price factors such as limits, observability, and portability.
Why model capability comes before input unit price
Native multimodal models are rarely billed on text tokens alone. Images, video, and files are encoded, chunked, or processed for visual understanding and then counted as input tokens. A larger context window lets a request carry more history, long documents, and media—but it can also raise input cost per call.
This article focuses on a large-parameter multimodal model with a relatively small activation footprint, long context, and support for text, image, video, and file input. For developers, parameter count matters less than four practical factors:
- How input tokens are counted
- Whether repeated content can hit cache
- The unit price of output tokens
- Whether batch or subscription capacity can lower unit cost
Comparing only “price per million input tokens” is incomplete. In customer support, knowledge-base Q&A, and code analysis, output tokens often dominate. In long-document Q&A, video analysis, and multi-turn chat, input volume and cache hit rate drive the bill.
Pay-as-you-go list prices: many channels look similar
Public pages for common aggregators and cloud channels often share the same base rates. Per million tokens, a typical standard input price is about CNY 0.8, cached input about CNY 0.23, and output about CNY 2.8.
Some first-party channels offer time-limited discounts at roughly half of standard:
- Input: CNY 0.4 per million tokens
- Cache hit: CNY 0.115 per million tokens
- Output: CNY 1.4 per million tokens
A limited-time discount is not a permanent price. Confirm validity window, eligible accounts, new-user requirements, and whether the discount applies to both cached input and output.
| Channel type | Input | Cache hit | Output | Notes |
|---|---|---|---|---|
| First-party, promotional | 0.4 CNY | 0.115 CNY | 1.4 CNY | ~50% off standard; watch expiry |
| First-party, standard | 0.8 CNY | 0.23 CNY | 2.8 CNY | Better for long-term steady usage |
| Cloud provider | 0.8 CNY | 0.23 CNY | 2.8 CNY | Often easier for procurement and ops |
| Aggregator | 0.8 CNY | 0.23 CNY | 2.8 CNY | Easy integration; check rate limits |
| Overseas-priced channel | ~1.04 CNY | ~0.21 CNY | ~3.45 CNY | FX and cross-border networking apply |
Overseas figures above use about CNY 6.9 per USD for estimation only. FX, taxes, payment fees, and settlement rules can change the actual amount.
Four factors that actually split cost
1. Time-limited discounts
For early development or low monthly volume, promotions are the simplest saving. Example: 10 million input tokens and 3 million output tokens per month.
Standard:
- Input: 10 × 0.8 = 8 CNY
- Output: 3 × 2.8 = 8.4 CNY
- Total ≈ 16.4 CNY
50% off:
- Input: 10 × 0.4 = 4 CNY
- Output: 3 × 1.4 = 4.2 CNY
- Total ≈ 8.2 CNY
Savings are modest at this volume but scale with usage. Teams should budget both promotional and post-promotion (standard) estimates so costs do not jump unexpectedly.
2. Cache hits
Caching fits repeated system prompts, fixed rules, long knowledge-base prefixes, and conversation history. Cached input is much cheaper than regular input, so high-repetition apps can cost far less than a naive input-total estimate.
Example: a Q&A system that always sends the same long policy document. If that prefix is cached, later requests need not pay full input price for it. Caching is not automatic savings. Watch:
- Whether the fixed prefix stays stable
- Enough overlap across requests
- Cache TTL and hit rules
- Dynamic content accidentally placed before the stable prefix
- Whether the SDK or gateway forwards cache-related fields
In practice, put stable instructions, rules, and knowledge summaries first; put user questions, timestamps, and live data last to improve reuse.
3. Batch APIs
Daily reports, archival, offline classification, bulk summarization, and data cleaning often do not need real-time responses. Batch trades latency for lower price.
Batch fits:
- High volume with low latency sensitivity
- Jobs that can finish in minutes or longer
- Many independent samples processed together
- Avoiding online concurrency quotas
If batch is priced at about 50% of standard, overall cost can approach a real-time promotional rate when input/output mix is unchanged. Batch is not suitable for live chat, real-time moderation, interactive coding, or instant decisions.
4. Subscription quotas and off-peak capacity
Some clouds sell prepaid token quotas or cheaper off-peak capacity. The point is not a lower public list price, but lower average cost via prepay, off-peak jobs, and capacity planning.
This fits:
- Relatively stable monthly usage
- Tolerance for delay
- Many offline jobs
- Teams that can schedule work centrally
If traffic is bursty and mostly peak-hour, prepaid quotas may not beat pay-as-you-go. Review one to three months of real usage, then model quota burn, expiry, off-peak share, and unused-quota risk.
Five typical usage patterns
Personal development and feature validation
Monthly volume is usually low. Prefer a channel with a clear promotion, solid docs, and simple key management. Do not add complex scheduling for tiny unit-price gaps.
Online support and knowledge-base Q&A
Repetition is high—optimize cache hit rate. Also watch output tokens: long answers raise cost. Cap max output, use structured reply templates, and layered retrieval.
Multimedia analysis
Images, video, and files inflate input. Preprocess before the model: compress unused images, extract key frames, drop duplicate pages, prefer document body text. The goal is less wasted input within product quality limits, not blindly lowering quality.
Batch content production
If instant return is not required, evaluate batch pricing. For titles, summaries, tags, and classification, submit in bulk and use retries, validation, and sharding for reliability.
High-concurrency production
Unit price is not enough. Check concurrency caps, tokens per minute, regional network quality, failover, and bill observability. A primary plus backup channel, with app-level wrapping of requests, retries, rate limits, and cost metrics, is often warranted.
Cost formula
Estimate monthly cost as:
monthly cost =input tokens / 1,000,000 × input unit price+ cached tokens / 1,000,000 × cache unit price+ output tokens / 1,000,000 × output unit price+ fixed capacity fees+ network, storage, and other add-ons
Count regular input and cached input separately. For multimodal work, also log image count, video duration, and file pages so you can locate cost spikes.
Non-price factors in vendor choice
The cheapest unit price is not always the lowest total cost. Also evaluate:
- Auth model and key permissions
- Request logs and audit
- Quota alerts and invoice detail
- Concurrency and rate-limit policy
- Regions and network stability
- Data compliance and privacy
- Standard APIs for migration
- Fast failover to a backup channel
For long-lived apps, isolate the model client: message format, timeouts, retries, monitoring, and cost stats. Price or vendor changes then do not force large business-logic rewrites.
Conclusion: fit the workload, then rank prices
Public base pay-as-you-go rates across many domestic channels are close. Real differences come from promotions, cache hits, batch, and subscriptions. Low-volume developers can start with discounts; high-repetition apps should invest in cache layout; offline bulk jobs fit batch; large steady traffic needs a combined view of quotas and off-peak capacity.
A sound decision order:
- Measure real input, cache, and output tokens
- Separate real-time from offline work
- Compute both standard and promotional bills
- Weigh concurrency, reliability, compliance, and migration cost
- Add monitoring and budget alerts
- Then pick the channel that fits the workload
Reliable cost conclusions come from model capability, call patterns, and the full bill—not a single input price or a short-lived discount.
