How LLM API pricing works: per token, per request and cached input
LLM APIs bill three ways — input and output tokens priced separately, cached input at a lower rate, and per request for image models — and the same model costs different amounts in different groups. Worked examples at live prices.
LLM APIs are priced three ways. Chat and embedding models bill per token, with input and output priced separately per million tokens. Input served from the context cache is billed at the cache-read price, which is usually much lower than normal input. Image generation and similar models bill a fixed price per request. Tutujin API balances are in CNY with no monthly fee, and the same model costs different amounts in different groups.
Three ways to bill
| Billing | Typical models | How it works | Unit |
|---|---|---|---|
| Per token | Chat, embeddings, some audio and image models | Input tokens × input price plus output tokens × output price | ¥ per 1M tokens |
| Cache read and write | Chat models with context caching | Cached input is billed at the cache-read price; some models also charge a cache-write price | ¥ per 1M tokens |
| Per request | Image generation, some search models | A fixed price per call, regardless of token count | ¥ per request |
Models and pricing shows which way each model is billed.
Reading a price table
Here is gpt-5.4-mini in every group:
| Group | Input ¥/1M | Output ¥/1M | Cache read ¥/1M |
|---|---|---|---|
| StandardLowest | ¥1.5 | ¥9 | ¥0.15 |
| Azure | ¥1.5 | ¥9 | ¥0.15 |
| Azure Plus | ¥2.25 | ¥13.5 | ¥0.225 |
| OpenAI Direct | ¥3.38 | ¥20.25 | ¥0.338 |
| OpenAI Direct Plus | ¥4.5 | ¥27 | ¥0.45 |
| Official · Enterprise | ¥8.21 | ¥49.27 | ¥0.821 |
gpt-5.4-mini · Prices in CNY, cheapest group for each model; updated Nov 12, 2026, 8:17 PM.
- Input: everything you send — system prompt, chat history, documents and tool definitions.
- Output: everything the model generates, including a reasoning model's thinking. Output usually costs several times as much as input.
- Cache read: the price for input served from the context cache; see What context caching is and how it saves money.
- Group: the same model is priced differently per group; see How to choose a group.
Cost of one request
For token-billed models:
cost = input tokens ÷ 1,000,000 × input price
+ cached tokens ÷ 1,000,000 × cache-read price
+ output tokens ÷ 1,000,000 × output priceSay a request has 3,000 input tokens and 800 output tokens. With deepseek-v4.1-flash:
Cost estimate · deepseek-v4.1-flash (China Models)
- Unit price
- input ¥2.8 / output ¥11.2 (per 1M tokens)
- Usage per request
- 3,000 input tokens, 800 output tokens
- Requests
- 1
- Cost per request
- ¥0.0174
Total about ¥0.0174Prices in CNY, cheapest group for each model; updated Nov 12, 2026, 8:17 PM.
The same request on Claude Sonnet 5.5:
Cost estimate · Claude Sonnet 5.5 (Claude Lite)
- Unit price
- input ¥3.2 / output ¥16 (per 1M tokens)
- Usage per request
- 3,000 input tokens, 800 output tokens
- Requests
- 1
- Cost per request
- ¥0.0224
Total about ¥0.0224Prices in CNY, cheapest group for each model; updated Nov 12, 2026, 8:17 PM.
The gap per request looks small; multiply by a month of traffic and it is not.
Cost of a month
Monthly cost = cost per request × number of requests. The same request 50,000 times a month:
Cost estimate · deepseek-v4.1-flash (China Models)
- Unit price
- input ¥2.8 / output ¥11.2 (per 1M tokens)
- Usage per request
- 3,000 input tokens, 800 output tokens
- Requests
- 50,000
- Cost per request
- ¥0.0174
Total about ¥868Prices in CNY, cheapest group for each model; updated Nov 12, 2026, 8:17 PM.
For a budget, take about 100 real requests, average the input and output tokens from each response's usage field, then multiply by expected volume. Guessed token counts are often far off.
Per-request models
Image models charge per call, however long the prompt:
| Model | Vendor | Cheapest group | Input ¥/1M | Output ¥/1M | Per request ¥ |
|---|---|---|---|---|---|
| gpt-image-2 | OpenAI | Standard | — | — | ¥0.12 |
| nano-banana-pro | Standard | — | — | ¥0.4 |
Prices in CNY, cheapest group for each model; updated Nov 12, 2026, 8:17 PM.
For example, 1,000 images:
Cost estimate · gpt-image-2 (Standard)
- Unit price
- ¥0.12 / request
- Requests
- 1,000
- Cost per request
- ¥0.12
Total about ¥120Prices in CNY, cheapest group for each model; updated Nov 12, 2026, 8:17 PM.
Four things that drive the bill
| Factor | Effect | What to do |
|---|---|---|
| Model tier | Unit prices can differ by tens of times between tiers | Use small models for simple tasks — see How to choose a model |
| Input length | Long prompts, history and documents are all billed as input | Trim prompts and history, use caching |
| Output length | Output is the expensive side, and thinking counts as output | Set max_tokens, ask for concise answers |
| Group | The same model is priced differently per group | Pick a group by how much stability you need |
More ideas in How to control API costs.
Checking what you were charged
- Each response's
usagefield returns input and output tokens, plus cached tokens when the cache was hit. - The Usage page in the console lists model, group, tokens and cost for every request.
- Billing rules in detail: Billing and usage.
FAQ
Is there a monthly fee or minimum spend?
No. Usage is deducted from your balance, which is in CNY, and you pay only for what you use.
Why does the same model cost different amounts in different groups?
Each group is a different line with different costs, so prices differ. You pick a group for each API key when you create it, and every request on that key is billed at that group's price.
Do I pay for a reasoning model's thinking?
Yes. Thinking counts as output tokens and is billed at the output price, so a short final answer can still cost a fair amount. Cap it with max_completion_tokens.
Can prices change?
They can. The models and pricing page shows when the price list was last updated, each request is billed at the price in effect at that time, and every charge is listed under Usage in the console.
Can I get an invoice?
We can issue overseas invoices; add our WeChat support JsnonoChat and send your order number. Chinese domestic invoices can be discussed case by case.