OpenAI-Compatible API Providers in China: A Practical Comparison (2026)

If you're building on the OpenAI SDK and want access to Chinese LLMs — Qwen, DeepSeek, GLM, MiniMax — you don't need a new SDK. Six major providers in China expose OpenAI-compatible /v1/chat/completions endpoints. Point your base_url at any of them, swap the model name, and your existing code works.

We compared all six across endpoint setup, pricing, model coverage, rate limits, and free tiers. This post is the reference we wish existed when we started routing requests through them.

TL;DR — Which Provider, in 30 Seconds

Provider Comparison Table

Endpoint Setup

Every provider follows the same pattern: set base_url, set api_key, pick a model. Here's the configuration for each:

Code: One SDK, Six Providers

The beauty of OpenAI-compatible APIs is that switching providers is a one-line change. Here's the same call against each provider using the OpenAI Python SDK:

from openai import OpenAI

# SiliconFlow
client = OpenAI(
    base_url="https://api.siliconflow.cn/v1",
    api_key="your-siliconflow-key",
)
response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V3",
    messages=[{"role": "user", "content": "Explain transformer attention in one paragraph."}],
)

# To switch to DashScope, change two lines:
client = OpenAI(
    base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
    api_key="your-dashscope-key",
)
response = client.chat.completions.create(
    model="qwen3.7-plus",
    messages=[{"role": "user", "content": "Explain transformer attention in one paragraph."}],
)

# DeepSeek direct:
client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key="your-deepseek-key",
)
response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Explain transformer attention in one paragraph."}],
)

The pattern is identical for Volcengine, Zhipu, and MiniMax — only base_url, api_key, and model change.

Pricing Comparison

Prices below are per million tokens in CNY (¥), sourced from each provider's official pricing page as of June 2026. For providers with tiered pricing (different rates for different context lengths), we list the base tier.

Key takeaway: DeepSeek direct offers the lowest per-token cost for both flash and flagship tiers. DashScope provides the broadest Qwen model selection with built-in tools. SiliconFlow offers multi-vendor model access on a single API key. MiniMax's M3 delivers frontier-class 1M-context reasoning at competitive rates.

Source note: Prices retrieved from official pricing pages on June 4, 2026. All providers adjust pricing periodically — verify current rates before making procurement decisions.

Rate Limits and Free Tiers

Decision Tree: Pick Your Provider

Pick DeepSeek direct if:

• You want the absolute lowest per-token cost
• DeepSeek models (V4 Flash, V4 Pro) cover your use case
• You don't need built-in tools or multi-model access on one key
• Periodic availability fluctuations are acceptable for your SLA

Pick DashScope if:

• You need official Qwen models (Qwen3.7-max, Qwen3.7-plus, Qwen3.6-flash)
• You want built-in tools (web search, code interpreter) without custom function calling
• You're already in the Alibaba Cloud ecosystem (unified billing, IAM, VPC endpoints)
• You need enterprise SLA and domestic compliance

Pick SiliconFlow if:

• You want one API key for models from DeepSeek, Qwen, GLM, MiniMax, and others
• You're prototyping and want to test multiple models quickly
• Free-tier models matter for development
• You want a single billing relationship for multi-vendor model access

Pick Volcengine Doubao if:

• You need Doubao models (ByteDance's proprietary LLM family)
• You need guaranteed TPM (tokens per minute) with SLA-backed throughput packages
• You're in the ByteDance/Volcengine ecosystem

Pick Zhipu if:

• You specifically need GLM-5 or GLM-5.1
• You want both OpenAI and Anthropic SDK compatibility from the same provider
• Coding Plan pricing fits your workload

Pick MiniMax if:

• You need 1M-context agentic reasoning (M3)
• Both OpenAI and Anthropic SDK compatibility matter
• You need speech, video, and image APIs alongside LLM

Routing Through Multiple Providers

If you don't want to pick just one, a routing layer can distribute requests across providers based on model, cost, or availability. TheRouter routes OpenAI-compatible requests through configured providers and supports provider/model routing and fallback — so you can set SiliconFlow as your primary for DeepSeek models and DashScope as your primary for Qwen, with automatic failover between them.

# Example TheRouter config (simplified)
providers:
  siliconflow:
    base_url: https://api.siliconflow.cn/v1
    api_key: ${SILICONFLOW_API_KEY}
  dashscope:
    base_url: https://dashscope.aliyuncs.com/compatible-mode/v1
    api_key: ${DASHSCOPE_API_KEY}
  deepseek:
    base_url: https://api.deepseek.com
    api_key: ${DEEPSEEK_API_KEY}

routes:
  - model: "deepseek-*"
    providers: [deepseek, siliconflow]  # fallback order
  - model: "qwen*"
    providers: [dashscope, siliconflow]

FAQ

Can I use the standard OpenAI Python/Node SDK with these providers?

Yes. All six providers implement the OpenAI-compatible /v1/chat/completions endpoint. Set base_url to the provider's endpoint and api_key to your provider-specific key. No SDK modification needed. See SiliconFlow setup or DashScope setup for step-by-step guides.

Which provider has the cheapest DeepSeek V4 access?

DeepSeek direct at ¥1/¥2 per million tokens for V4 Flash (input/output). SiliconFlow also offers DeepSeek models at similar rates. DeepSeek's automatic context caching can reduce effective input costs by up to 98% for repetitive prompts.

Are these providers available outside China?

DashScope has international regions (Singapore, US, Frankfurt). DeepSeek's API is accessible globally. SiliconFlow serves international users at api.siliconflow.cn. MiniMax operates platform.minimax.io for international access. Volcengine and Zhipu are primarily domestic, though API access may work from outside China with added latency.

Do these providers support function calling and tool use?

All six support function calling via the OpenAI-compatible API format. DashScope additionally offers built-in tools (web search, code interpreter, web scraping) that don't require custom function definitions. Feature support varies by model — check each provider's documentation for model-specific capabilities.

Can I use these providers with coding agents like Claude Code or Cursor?

Yes. DeepSeek explicitly documents agent integration guides for Claude Code, GitHub Copilot, and OpenCode. Any OpenAI-compatible provider can work with tools that accept a custom base_url. See our model pages for provider-specific setup notes.

What about streaming and structured output support?

All six providers support streaming responses via stream: true. Structured output (JSON mode) support varies: DashScope's Qwen models and Zhipu's GLM models have the broadest structured output support. DeepSeek and MiniMax support JSON mode for most models. Check each provider's model documentation for specific capabilities.

---

Prices and model availability retrieved from official provider documentation on June 4, 2026. This post compares upstream API providers — TheRouter is one way to route between them. Provider pricing changes frequently; verify current rates on each provider's pricing page before making decisions.