Published at: 2026-09-17
Billing & usage management admin guide
This guide is for system administrators who plan AI resource budgets and control costs.
Credit consumption estimation methods
Agent execution estimation
Total credits for one Agent execution (one question and one answer) = intent recognition consumption + all prompt template consumption + all AI operation consumption + final content integration output consumption.
Credits per stage = input character count x character-to-token conversion ratio (1.2 for Chinese, 1.3 for English).
| Stage | Description |
|---|---|
| Intent recognition | The input is usually the command a user enters in the Agent input box, which rarely exceeds 200 characters. The output is the Skill and operation the Agent will execute, which is also small. Estimate it at about 1000 credits |
| Final content integration | The input is usually the sum of all operation outputs during Agent execution. You can estimate its character count directly. The output is usually the content of the user command |
| Prompt template | The input is the content expected to fill the scenario variables configured in the prompt template. You can estimate the character count based on the variable meaning. The output is the content defined in the template |
| AI operation | First break down how many AI operations a user command will execute, then estimate based on the complexity reference range |
AI operation consumption reference:
| AI operation | Estimated consumption (credits per call) |
|---|---|
| Web search | 0.5-1 |
| Create object data operation (create account, schedule, and so on) | 0.12-0.5 |
| Query object data operation | 0.3-1 |
| Knowledge base retrieval operation | 0.2-0.4 |
| Text parsing to text | Dynamically uses a parsing model based on file type and document resolution |
Flow estimation
Based on the Flow design, identify the AI operations and prompt templates used in the Flow, and estimate the credit consumption for each node. The billing reference for prompt templates and AI operations in a Flow is the same as the table above. It uses the same billing logic as when used in an Agent.
RAG knowledge chunk vectorization ingestion
When you use a knowledge base for intelligent Q&A or retrieval to generate content, you need to consider the cost of vectorizing and ingesting knowledge chunks. You also need to consider the cost of recalling relevant knowledge chunks based on the user question.
Chunk vectorization ingestion reference:
- Document: PDF file 0.1 credits per page.
- Image: 0.1 credits per image.
Scenario examples
The following consumption values are averages from normal use. They are for reference only. Actual consumption may vary based on user requirements and task complexity.
| Scenario | Example case | Model used | Estimated credit consumption |
|---|---|---|---|
| Daily customer service conversation | One question and one answer. The AI Customer Service retrieves knowledge base content and answers the question | deepseek-v3.2 | 0.3 credits |
| Long text summary | Input 50,000 characters, output 2000 characters | deepseek-v3 | 1.4 credits |
| Account intelligence gathering and insight | Gathers Business Information, CRM information, and web search information to provide all-around insight on an account or enterprise | deepseek-v3.2 | 1.2 credits |
| Weekly and daily report summary | Summarizes 10 sales records from one week | deepseek-v3.2 | 0.1 credits |
| Content quality inspection | Inspects the content quality of sales records, field visits, and opportunity follow-ups filled in by sales. About 300 characters input each time | deepseek-v3 | 0.06 credits |
| Knowledge base Q&A | Intelligent Q&A based on a knowledge base | deepseek-v3 | 0.4 credits per call |
| File parsing | Converts a file to text and summarizes it on request. Normal precision required. The file has 10 pages, 20,000 characters, and outputs 200 characters of key information | deepseek-v3.2 | About 0.8 credits per call |
Cost control recommendations
| Method | Description |
|---|---|
| Choose the right model | Use a low-cost model for simple tasks and a high-end model for complex tasks |
| Control context length | Limit the number of conversation history turns to avoid unbounded context growth |
| Optimize prompts and instructions | Streamline input content to reduce invalid tokens |
| Batch processing | Merge multiple small tasks to reduce the number of calls |
Rapid consumption troubleshooting
| Cause of rapid consumption | Details | Solution |
|---|---|---|
| Using high-end models | High-end models such as Deepseek-r1 and Doubao seed2.0 pro are expensive | Run an A/B test to evaluate whether a high-end model is necessary. Evaluate whether an economy model achieves the same result |
| Long context | Conversation history accumulates, which increases input tokens per turn | Reset conversations regularly to control context length |
| Frequent calls | High-frequency calls increase cumulative consumption | Merge tasks to reduce the number of calls |
| Multi-turn reasoning loops | Even simple tasks can trigger multiple “think act observe” loops. Each turn injects a large number of prompts and tool definitions | Streamline tool mounting and keep only the tools the current task requires. Also optimize instructions based on task complexity to avoid unnecessary action planning and execution by the AI |