Pricing and Billing Models for Native Multimodal LLM APIs: A Cross-Channel Comparison

    This article breaks down how native multimodal large language model APIs are priced across official platforms, domestic cloud providers, and international channels. It compares input, cached-input, and output token rates, walks through five typical usage scenarios with concrete cost calculations, and explains how to build a realistic cost model so you can choose the right billing mode—on-demand, batch, or off-peak—based on your workload rather than headline unit prices alone.

    1. The Short Version: Unit Prices Are Similar, Billing Mechanics Determine Your Real Cost

    For any given large model, developers usually track three price metrics: input token price, cached input token price, and output token price. Based on publicly available information, the standard list prices across major channels are largely aligned:

    Channel Input Price Cached Input Output Price Key Characteristics
    Official platform, limited-time pricing ¥0.4 / 1M tokens ¥0.115 / 1M tokens ¥1.4 / 1M tokens 50% off standard price; usually time-limited
    Official platform, standard pricing ¥0.8 / 1M tokens ¥0.23 / 1M tokens ¥2.8 / 1M tokens Suitable as a long-term cost baseline
    Domestic cloud provider A ¥0.8 / 1M tokens ¥0.23 / 1M tokens ¥2.8 / 1M tokens Typically offers an OpenAI-compatible endpoint
    Domestic cloud provider B ¥0.8 / 1M tokens ¥0.23 / 1M tokens ¥2.8 / 1M tokens Prices may vary by region
    Domestic cloud provider C ¥0.8 / 1M tokens ¥0.23 / 1M tokens ¥2.8 / 1M tokens Convenient for managing multiple models in one place
    International channel ~¥1.04 / 1M tokens ~¥0.21 / 1M tokens ~¥3.45 / 1M tokens Priced in USD; affected by exchange rates
    International aggregator ~¥1.04 / 1M tokens ~¥0.21 / 1M tokens ~¥3.45 / 1M tokens Easy access to multiple models, but overall cost tends to be higher

    All prices are per million tokens and are shown for cost comparison only. Actual charges may also be affected by region, taxes, account balance, concurrency limits, promotional periods, and settlement rules.

    On the surface, there is little difference in base unit prices between domestic channels. The four factors that actually drive total cost are:

    1. Whether limited-time discounts apply;
    2. Whether requests are suitable for batch asynchronous processing;
    3. Whether a high cache hit rate can be achieved;
    4. Whether off-peak execution or subscription/credit-based plans are acceptable.

    In other words, comparing only “how much does input cost and how much does output cost” is not enough to determine which option is cheapest.

    2. Understanding the Price Structure

    2.1 Input Token Pricing

    Input tokens include user questions, system prompts, conversation history, tool-call parameters, and the encoded representations of multimodal content. For text tasks, input token counts are relatively easy to estimate. For images, video, and files, however, actual consumption depends on resolution, frame count, page count, content complexity, and how the platform preprocesses the input.

    For example, a single document Q&A request may include:

    • A fixed system instruction;
    • Business rules and field definitions;
    • The content of uploaded files;
    • Multi-turn conversation history;
    • The current question.

    If a large amount of fixed content is resent on every request, input costs will climb quickly. Engineering practices that help reduce consumption include shortening system prompts, compressing conversation history, splitting long documents, and enabling context caching.

    2.2 Cached Input Token Pricing

    Cache hits typically apply to repeated context prefixes, such as fixed system prompts, long-lived business knowledge, consistent formatting constraints, and identical tool definitions.

    The cached rate is significantly lower than the standard input rate, but that does not mean all repeated text automatically qualifies for the discount. Whether a cache hit actually occurs usually depends on:

    • Whether the repeated content sits at the beginning of the context;
    • Whether the interval between requests exceeds the cache’s time-to-live;
    • Whether the text has changed even slightly;
    • Whether the platform enforces a minimum length for caching;
    • Whether multimodal content is eligible for caching;
    • Whether different model versions or service regions share the same cache.

    So when estimating cache savings, do not assume “all history will hit the cache.” Measure your actual hit rate from request logs instead.

    2.3 Output Token Pricing

    Output tokens are usually more expensive than input tokens, and this gap is most visible in long-form generation, structured results, code generation, report writing, and video analysis summaries. If your application tends to produce verbose answers, constrain the output format at the prompt level—for example, ask the model to return only required fields, cap the response length, or use structured data instead of natural-language descriptions.

    Note that setting a maximum output length only prevents abnormally long replies; it does not guarantee shorter responses. A more robust approach is to combine output format design with result validation and timeout/retry mechanisms.

    3. Cost Comparison Across Five Typical Scenarios

    Scenario 1: Low-Frequency Personal Projects

    Assume 10 million input tokens and 3 million output tokens per month, with essentially no cache hits:

    • Standard pricing: 10M × ¥0.8/1M + 3M × ¥2.8/1M = ¥16.4;
    • Limited-time 50% discount: 10M × ¥0.4/1M + 3M × ¥1.4/1M = ¥8.2.

    For low-frequency users, absolute cost is not the main concern—focus instead on integration stability, API compatibility, and ease of debugging. If the discounted pricing remains available for a long time, the official discounted channel usually wins. But if the promotion has an expiration date, evaluate the project’s long-term viability at standard prices.

    Scenario 2: Real-Time Conversational Applications

    Real-time customer support, Q&A bots, and interactive assistants require low latency, so on-demand invocation is the natural fit. Cost characteristics of this workload: high request volume, high context repetition, and variable output length.

    Optimization priorities include:

    • Placing stable role definitions in a fixed prefix;
    • Trimming meaningless conversation history;
    • Summarizing overlong sessions;
    • Setting a sensible maximum output length;
    • Caching answers to repeated questions at the application layer;
    • Applying rate limits to anomalous requests.

    With a high cache hit rate, actual costs can be well below estimates based on standard input pricing. Conversely, if the full context changes every turn, the cache discount rarely helps.

    Scenario 3: Batch Document Processing

    Contract parsing, invoice extraction, knowledge base cleanup, and archival of historical documents generally do not require immediate responses, making batch APIs or asynchronous tasks a better fit. Under publicly documented rules, batch processing is often billed at a discount to the standard price—in the examples used here, batch pricing is 50% of standard.

    Take 100 million input tokens and 20 million output tokens:

    • Standard real-time calls: 100 × 0.8 + 20 × 2.8 = ¥136;
    • Batch at 50%: ¥136 × 50% = ¥68.

    The advantages of batch mode go beyond pricing: it simplifies scheduling, retry handling, and persisting results. However, it is unsuitable for interactive workloads where users are waiting. Enterprises should separate real-time tasks from offline ones rather than sacrificing user experience to save on unit price.

    Scenario 4: Large-Scale Production Workloads

    At monthly volumes in the billions of tokens, small unit-price differences get amplified. For example, with 2 billion input tokens and 500 million output tokens at standard pricing:

    • Input cost: 2000 × 0.8 = ¥1,600;
    • Output cost: 500 × 2.8 = ¥1,400;
    • Total: ¥3,000.

    At a 50% discount, the theoretical cost drops to ¥1,500. Each additional month the discount stays in effect yields meaningful savings—but don’t evaluate promotions on price alone. Check whether the promotion has an expiration date, whether it restricts account types, concurrency, or regions, and whether you can migrate smoothly if pricing changes.

    Production environments also demand attention to quotas, rate limiting, failover, billing detail, and API key permission management. A cheap endpoint that is unreliable can ultimately cause far greater business losses.

    Scenario 5: Off-Peak-Tolerant Workloads

    Some channels offer subscription credits, prepaid quotas, or off-peak resource discounts. For example, certain subscription models may bill off-peak calls at 30% of the regular price. For overnight batch jobs, offline data cleaning, and periodic report generation, such plans can be cheaper than standard pay-as-you-go.

    That said, off-peak pricing usually comes with scheduling uncertainty:

    • Longer queue times;
    • Unpredictable completion times;
    • Insufficient capacity during peak periods;
    • Unsuitability for strictly real-time workloads;
    • Expiration dates or usage restrictions on subscription credits.

    Off-peak plans are therefore best suited to tasks that are retryable, pausable, and delay-tolerant—not online transactions or real-time customer support.

    4. Building Your Own Cost Model

    Use this general formula to estimate monthly costs:

    1. Monthly cost =
    2. standard input tokens × standard input price
    3. + cached input tokens × cached price
    4. + output tokens × output price
    5. + other fixed fees

    If batch discounts apply, apply a discount coefficient to the base cost:

    1. Batch cost = base cost × batch discount coefficient

    Maintain at least three budgets:

    1. Conservative budget: everything at standard prices, with no assumed cache hits or discounts;
    2. Realistic budget: based on measured cache hit rates and average output length;
    3. Optimistic budget: factoring in promotional pricing, batch discounts, or off-peak discounts.

    This keeps your project from over-relying on short-term promotions. For enterprise applications, also include retries, failed requests, evaluation runs, and development/testing requests in the budget—otherwise the post-launch bill will almost certainly exceed theoretical estimates.

    5. Price Is Not the Only Criterion When Choosing a Channel

    Reliability

    Evaluate service availability, failure recovery, request timeouts, and rate-limiting policies. For critical workloads, design a fallback channel or a downgrade path to a smaller model—but keep in mind that request formats and response fields may differ between channels.

    Compatibility

    Channels that support a common API protocol are usually easier to migrate to. But protocol compatibility does not mean full parity: still verify multimodal message formats, file upload methods, streaming output, tool calling, and error-code handling.

    Data and Compliance

    When handling contracts, financial data, medical data, or internal business data, confirm the data storage region, log retention period, policy on training usage, and access control capabilities. Never let a low price override data security requirements.

    Billing Transparency

    Mature cost management should be able to distinguish input tokens, cached input, output tokens, batch jobs, and failed retries. Only with sufficiently granular billing can a development team pinpoint cost anomalies.

    6. Final Recommendations

    • Personal projects, prototyping, or low-frequency calls: prioritize pay-as-you-go channels with transparent pricing and simple onboarding.
    • Real-time conversational workloads: compare latency, reliability, caching mechanisms, and concurrency capacity—not just output unit price.
    • Batch document processing, data cleaning, and periodic report generation: prioritize batch discounts or off-peak resources.
    • High monthly volume: build a long-term budget on standard prices and treat limited-time discounts as upside, not as a permanent price.

    In summary, list prices for the same model can be nearly identical across platforms. What actually determines your final spend is your invocation pattern, cache hit rate, real-time requirements, and how long discounts last. The sensible approach is not simply hunting for the lowest unit price, but splitting tasks by workload type first, choosing an appropriate billing mode for real-time, batch, and off-peak tasks respectively, and then continuously calibrating your cost model with logs and billing data.