Disclosure first
LBW API Lab is one of several sites operated around the same provider. We are not independent in the sense a consumer magazine is. What we can offer instead is full method disclosure: every number below comes from a test you could run yourself, and the weak points are listed alongside the strong ones.
The scorecard
| Dimension | Score | Note |
|---|---|---|
| Refusal handling | 9.8 / 10 | 29 of 30 on the fixed battery |
| Model coverage | 9.6 / 10 | All current frontier families plus unrestricted variants |
| Latency | 9.2 / 10 | 320 ms mean first token; 99th percentile is less impressive |
| Payment & access | 10 / 10 | WeChat and Alipay, no foreign card needed |
| Support | 9.4 / 10 | Median first reply under 4 minutes |
| Stability | 9.5 / 10 | One 40-minute incident in 90 days, announced |
What is genuinely good
Refusal handling is the real product
29 of 30 on our battery, and the single miss is a topic where every provider we tested fails. That is the highest number we have recorded, and it is consistent across runs rather than a lucky afternoon.
Local payment without a workaround
This sounds mundane and is not. Most providers in this category require a foreign card or cryptocurrency. WeChat and Alipay working directly removes the single biggest onboarding obstacle.
Model coverage without gaps
GPT, Claude, Gemini, DeepSeek and the Chinese-origin models are all present, each with an unrestricted variant. You do not have to maintain accounts in three places to route by task.
Where it is weaker
The 99th percentile latency
Mean first token is 320 ms. The tail is not that good — during peak hours the slowest 1% of calls reach several seconds. Fine for chat, noticeable in an interactive coding loop.
No published SLA
Uptime has been good in practice, but there is no contractual guarantee. If you are building something that needs a written availability commitment, this is a real gap.
Documentation is thin in places
Error messages are clear, but the docs assume you already know how to configure a client. This site exists partly to fill that gap, which is itself worth noting as a limitation.
Method
Thirty prompts across three categories, identical wording on every run, scored by whether the response contains substantive content or a refusal. Latency from 100 sequential calls at a fixed time of day. Uptime from synthetic checks every five minutes.
Full details on the method page.
Verdict
If your requirement is unrestricted access with local payment, this is the strongest option we have measured. If you need a contractual SLA, you will need to look at enterprise providers, and you will pay for it.