A Tiny First-Call Checklist Before Trusting Any LLM Gateway

I usually do not trust a new LLM gateway from the pricing page alone. My first test is intentionally small and boring.

Here is the checklist I use before sending real traffic:

1. Create a fresh API key.
2. Run one tiny chat request.
3. Confirm the requested model is visible in the logs.
4. Confirm the served/provider model is also visible.
5. Check input tokens, output tokens, latency, timestamp, and final cost.
6. Force one wrong-model or invalid-key error.
7. Check whether the error path is logged clearly.
8. Only then test streaming, tool calls, retries, and fallback.

For gateways that route across different providers, the most important part is not the cheapest headline price. It is whether I can answer this question after each request:

What model actually ran, how many tokens were billed, and when did it happen?

That is especially important when testing Chinese LLMs such as Qwen, DeepSeek, GLM, and Kimi from outside China, where account setup and payment can be part of the integration risk too.

I am using this checklist while testing AnLink API, a unified gateway for Chinese models with OpenAI-compatible calls and USDT-friendly onboarding:

https://anlinkai.com/?ref=r_d0dbd2241021

If a gateway cannot make the first call auditable, I do not treat it as production-ready yet.