Low minimum top-ups and generous rate discounts, suited to individual developers and small-scale testing. Price and stability usually trade off against each other — start with a small top-up.
An AI API relay hosted in mainland China and reachable without a VPN. It covers the Claude, GPT…
This category is aimed at individual developers on a budget and at small-scale testing. The defining features are a low minimum top-up and generous rates.
One thing needs saying up front: a low price is not a problem in itself, but a low price usually comes with trade-offs. The usual ones are a longer path (more layers of forwarding), lower concurrency limits, downgrading on expensive models, or simply a short-term promotion to pull traffic in. None of these is inevitable, but you should know what to check for.
The pragmatic approach is a small test followed by staged scaling: top up the minimum, run real tasks for one to two weeks, and watch three things — whether the error rate is clearly higher than elsewhere, whether output quality matches the official channel, and whether the charges line up with expectations. Only scale up once all three look right.
And do not read the rate in isolation. A rate is a coefficient against the official price, but billing also depends on whether input and output are priced separately, whether caching is charged, and how a reasoning model's thinking tokens are counted. The same rate under different billing rules can produce bills that differ several times over.
The picks below are computed live from the current measurements, not fixed conclusions — change the data and the picks change with it. Each states its basis, and you can open the detail page to check it yourself.
The top 1 services by overall score in this category. 1 of them are confirmed to expose a /v1/models endpoint — that figure comes from probing, not from the services' own claims.
| Service | Region | Uptime | Latency | Models | Payment | API |
|---|---|---|---|---|---|---|
|
Lever AI
leverai.net
|
China direct | 98.36% | 5665 ms | — | Alipay · WeChat Pay | Has /v1/models |
The questions we see most often about this category. The guidance applies to the category as a whole, not to any single service.
You cannot convert it directly. A rate is a coefficient against some baseline, and services pick different baselines; their billing rules differ too (whether input and output are split, whether cache hits are charged). The reliable method is to run your own traffic for a while and divide the actual bill by the token count to work backwards.
Our article on the six common patterns among cheap relay services breaks them down in detail. The main types: a lightweight model passed off as a flagship; quality maintained only for requests of a certain length; traffic shifted to a cheap channel at peak times; and a sharp price increase after a short promotion. What they share is that every surface metric still looks normal — you have to judge by actual output quality.
Start with the minimum top-up and run one to two weeks of real work before adding more. The first day or two is not enough to conclude anything — some problems only surface once volume rises, such as concurrency limits, or quality dropping after a free allowance runs out.
Not as your only channel. A sensible use is as a standby or a cost-offload path, with the primary traffic on a service you can verify and that offers some commitment. If you do rely on one, keep at least one alternative configured so a single failure does not take the business down.
It depends on your workload. Batch jobs (offline data labelling, bulk document processing) are insensitive to latency, so a low rate is worth more there; interactive applications (code completion, customer-support chat) are latency-sensitive, and the money saved may not cover the loss in experience. Our comparison board sorts by latency, so you can filter by scenario.
This page shows you what exists; the pages below show you who is more stable, faster and cheaper. Read together, they tell you more than either alone.