Skip to main content
OpenPAI is an API gateway: your app → OpenPAI → the upstream model provider. Understanding a few key concepts before you start will save you some pitfalls.

​
Account (User)

Represents a real user or team. An account holds quota, and every API call is billed against it. An account belongs to a group, which determines the multiplier applied when the account uses a model.

​
Token (API key)

A string like sk-XXXXXXXX that serves as your credential when calling the API. One account can create multiple tokens, and each token can be configured separately:
  • Quota limit: the maximum amount this token may spend (-1 means unlimited).
  • Expiry: leave empty for no expiry.
  • Model scope: an allowlist of models this token may call (empty = all).
  • IP allowlist: restrict calls to specific IPs (optional).
Create a separate token for each app / environment (test, production) to make tracking spend and revoking easier.

​
Group

Groups implement tiered pricing. For example: When different groups call the same model, the actual charge is multiplied by the group’s multiplier. Groups can also be linked to channel groups — higher tiers are routed to more stable upstream channels.

​
Channel

A channel is the connection config between OpenPAI and an upstream model service. One model can be served by multiple channels:
Each request randomly selects a healthy channel by weight, and automatically retries the next channel on failure. Channel-level details are completely transparent to the caller — you only need the model ID.

​
Model

Model IDs match the upstream’s official names, e.g. gpt-5.5, claude-opus-4-8, gemini-3.1-pro-preview. In the OpenPAI console or the Models overview you can view:
  • Model ID and its upstream
  • Input / output price multipliers
  • Context window
  • Supported capabilities (vision, tools, streaming, caching, etc.)

​
Tokens and quota

OpenPAI uses token-based billing, consistent with OpenAI:
  • prompt_tokens: the input tokens you send to the model.
  • completion_tokens: the output tokens the model returns.
  • cached_tokens: tokens that hit the prompt cache, billed at a lower rate.
  • reasoning_tokens: chain-of-thought tokens produced by reasoning models (invisible but billed).
The final charge (simplified):
See Billing for the detailed rules.

​
Endpoint

OpenPAI exposes 3 protocol endpoints at once; pick the one that matches your existing code: All three endpoints share the same sk- key and account quota — just choose the protocol closest to your existing application.