Seven checks before you top up a Chinese LLM API gateway
Most people choose a gateway by price, then discover the expensive part is everything else: a card that gets reversed, a model build three releases behind, or an error message that tells you nothing at 2am.
- Does it take a rail you actually have?
"Credit cards accepted" is not "your card will work". Cross-border AI charges get reversed by issuer fraud engines after authorisation, and the merchant only sees a decline. Ask what else it takes: PayPal rides its own account instead of your bank's pipeline, and USDT removes the card network entirely.
- Prepaid or subscription?
A subscription adds a failure mode you cannot see. A renewal that quietly fails takes a working integration with it. A prepaid balance just runs down, and you can watch it run down.
- Which build do you actually get?
Model names are stable; builds are not. "V4.1 Flash" can mean any build since launch. Ask which build they serve, and whether they will tell you when it changes.
- Can you reproduce a failure?
Trigger a 402 on purpose before you commit, and read what comes back. A generic 500 for a billing problem is a provider you cannot debug.
- Does it speak the format your tools already use?
Claude Code, Cursor, Cline, Dify and the OpenAI SDKs all speak one of two formats. Moving gateways should be a base URL change, not a rewrite.
deepseek
api
webdev
- Are retries billed?
Ask whether a failed upstream call is retried before it reaches you, and whether you pay for the retry. A gateway that bills for retries is more expensive than its rate card.
- Who is behind it?
A support address that answers, and a page that says who operates the service, beat any benchmark.
I run one of these (NovaAPI), so take that with the usual salt. The checks above are the ones I would run on anyone.
Full version with the failure modes: https://novaapi.lbase.com/blog/choose-chinese-llm-api-gateway.html
Top comments (0)