Metering Guide
To help developers and enterprise users utilize AtomGit's model capabilities more efficiently, this metering guide provides a unified explanation of Token consumption, model types, and Notebook/Space core-hour calculations, helping you understand resource calculation rules and estimation methods clearly.
I. Overview of Metering Methods
AtomGit currently offers two types of resource metering methods:
- Token Metering: Suitable for model inference services such as text generation, image-to-text generation, text-to-image generation, sentence similarity, and automatic speech recognition.
- Core-Hour Metering: Suitable for scenarios requiring computing resources, such as Notebook and Space.
The metering methods for different services are calculated independently and do not affect each other.
II. Token Metering Rules
Different model types consume different numbers of Tokens. The following are the estimated Token rules for common model types:
| Model Type | Chinese Token Estimation Rule | English Token Estimation Rule | Supplementary Notes |
|---|---|---|---|
| Text Generation | 1 Token ≈ 1.5–1.8 Chinese Characters | 1 Token ≈ 4 English Characters | Different models use different tokenization algorithms, leading to slight variations in actual Token counts. |
| Image-to-Text | 1 Token ≈ 1.5–2 Chinese Characters | 1 Token ≈ 4 English Characters | Image-related Tokens are calculated based on resolution: e.g., 512×512 ≈ 334 Tokens. |
| Sentence Similarity | 1 Token ≈ 1.5–2 Chinese Characters | 1 Token ≈ 4 English Characters | The similarity score consumes approximately 4 Tokens. |
| Automatic Speech Recognition | 50,000 Tokens per session | Each successful generation of an audio file under 50M consumes 50,000 Tokens. | |
| Text-to-Image | 50,000 Tokens per image | Token consumption may vary with resolution and quality; refer to the actual call result. |
Note: The data in the table is a general estimation rule; actual Token consumption is based on the API response.
III. Core-Hour Calculation Rules (Notebook / Space)
Notebook and Space use CPU resources and are measured using core-hours.
Core-Hour Calculation Formula:
Core-Hours = Number of CPU Cores × Running Time (hours)
Reference Table for Notebook/Space Core-Hour Consumption
| CPU Cores | 10 Minutes | 30 Minutes | 60 Minutes | 120 Minutes |
|---|---|---|---|---|
| 0.5 Core | 0.08 | 0.25 | 0.5 | 1 |
| 2 Cores | 0.34 | 1 | 2 | 4 |
| 4 Cores | 0.67 | 2 | 4 | 8 |
| 8 Cores | 1.34 | 4 | 8 | 16 |
| 16 Cores | 3.34 | 8 | 16 | 32 |
| 32 Cores | 5.34 | 16 | 32 | 64 |
Notes:
- Notebook and Space have identical metering logic;
- Actual deduction is calculated based on resource usage logs, and the final charge is based on actual usage logs;
- Multiple concurrent instances are calculated by adding up each instance;
IV. Frequently Asked Questions
1. How is the total Token consumption calculated?
Total Tokens = Input Tokens + Output Tokens
The longer the input and output, the higher the total Token consumption.
2. How are mixed Chinese and English Tokens calculated?
The system automatically identifies based on the model's tokenization strategy; no manual distinction is needed for mixed Chinese and English.
3. Why does the image Token change?
It depends on the image size and the model's supported resolution; generally, the higher the resolution, the more Tokens are consumed.
This concludes the platform's metering method explanation. Whether you're using the inference API, the online model experience, or running code in Notebook/Space, you can estimate resource consumption according to the table rules. We hope this document helps you better understand the calculation methods for Tokens and core-hours, allowing you to use the platform's various models and services more confidently. If you have any questions, feel free to contact us anytime in the community.