Platform concepts
Model IDs, supply channels, routing, and fallback models.
Model IDs
Model IDs use developer/model-name, for example deepseek/deepseek-v4-pro and z-ai/glm-5.2. The prefix identifies the developer, such as DeepSeek or Zhipu, and the remaining part identifies the model. See the catalog for all models on sale, or call GET /v1/models for those available to the current key.
A model ID may have several supply channels. The platform selects a channel automatically based on latency and availability, transparently to the caller.
Supply channels
| Channel | Description |
|---|---|
| Platform-operated infrastructure | Open-source model weights deployed on our own Chinese accelerator clusters, with control over supply and pricing |
| Official access | Supply through model vendors’ official or partner channels |
See each model detail page for channel types, current availability, and latency. Data sources are described in the public platform API.
Intelligent routing and fallback
When a channel fails or rate-limits a request, the gateway switches to another channel for the same model.
You may also specify a request-level fallbacks list. If the primary model fails because of an upstream error, timeout, or rate limit, the gateway retries a fallback model with the same parameters. The response’s model field identifies the model actually used:
{
"model": "z-ai/glm-5.2",
"fallbacks": ["deepseek/deepseek-v4-pro"],
"messages": [{"role": "user", "content": "..."}]
}fallbacks extends the OpenAI protocol. Put it at the top level of the request body as shown above. SDKs that do not recognize extension parameters can pass it through extra_body.
Billing
Usage is charged per token (¥/million tokens), with separate input, output, and cache-read prices. Some models have peak and off-peak prices. See billing.