Switching From OpenAI to Chinese Models: A Drop-In Migration Guide

Because almost every Chinese provider exposes an OpenAI-compatible Chat Completions endpoint, moving code from OpenAI to DeepSeek, Qwen, GLM or Kimi is usually a configuration change rather than a rewrite. The real work is choosing the right model, checking which features you rely on, and validating quality before you switch production traffic.

The two-line change

If you use the official OpenAI SDK, the change is the base URL and the model id:

# before
client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

# after
client = OpenAI(
    base_url="https://aiapi-pro.com/v1",
    api_key=os.environ["NOVAI_API_KEY"],
)

resp = client.chat.completions.create(
    model="deepseek-v4-flash",   # was "gpt-4o-mini"
    messages=[{"role": "user", "content": "Hello"}],
)

Model ids, context windows and prices are listed at /model-marketplace.html and in the machine-readable feed at /china-ai-leaderboard.json. If you use a framework (LangChain, LlamaIndex, the Vercel AI SDK), it normally reads baseURL from configuration.

Model mapping: a starting point

These are reasonable starting points, not drop-in equivalents. Always re-run your own evaluation set.

Node / TypeScript

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://aiapi-pro.com/v1",
  apiKey: process.env.NOVAI_API_KEY,
});

const res = await client.chat.completions.create({
  model: "deepseek-v4-flash",
  messages: [{ role: "user", content: "Hello" }],
});
console.log(res.choices[0].message.content);

What breaks (and what to check first)

• Tool calling. Supported, but schema strictness and parallel-call behaviour vary by model. Re-test your tool loop, not just a single call.
• JSON mode / structured output. Some models support response_format, some only via prompting. Validate with your real schema.
• Vision. Only vision models accept image content; passing an image URL to a text-only model returns an error.
• Token accounting. Chinese and OpenAI tokenizers differ, so the same prompt has a different prompt_tokens count. Budget from measured usage, not estimates.
• OpenAI-only surface area. The Responses API, Assistants and some built-in tools are not part of the Chat Completions compatibility layer.
• Rate limits. Limits are per account/key and differ from your OpenAI tier; test at your real concurrency.

Feature matrix

A safe rollout plan

1. Build a 50–200 prompt evaluation set from your real traffic (including the ugly edge cases).
2. Shadow-compare the new model against your current one; score correctness, not just similarity.
3. Route a small % of traffic with a kill switch and a fallback model in the same client.
4. Measure cost per successful task rather than cost per token — retries and failures are part of the cost.
5. Watch latency by region; the fastest route to a provider depends on where your servers are.

For a worked cost comparison of Chinese models on the same task, see the coding-agent cost model and the measured-cost hub.

FAQ

Can I keep using the OpenAI Python SDK?

Yes. Set base_url to the OpenAI-compatible endpoint and pass a model id from the provider's catalogue. Chat Completions calls, streaming and tools work; OpenAI-only APIs such as Assistants and the Responses API do not.

Is the request/response format identical?

For Chat Completions it is compatible, but individual models differ in tool-calling strictness, JSON-mode support, vision input and tokenizer behaviour. Test your specific features.

Do I need to rewrite prompts?

Usually not, but Chinese models may respond differently to system prompts, formatting instructions and refusal policies. Run your evaluation set before switching production traffic.

How do I compare costs fairly?

Compare cost per successful task, not price per token: include retries, cache hit rates and output token distribution. A cheaper model that needs two attempts is often more expensive.

Can I mix models in one application?

Yes. Use one base URL and key, then choose per task: a cheap model for extraction and a stronger one for reasoning or coding. Keep a fallback model configured.