The 2026 relay site ecosystem in mainland China: three kinds of operator behind 2,000 sites

2,000 sites, but only three ways to survive

The number of relay sites in mainland China passed 2,000 in 2026. That sounds implausible until you look at the upstream: the vast majority are the same open-source project underneath. The industry saying is that more than 80% of commercial relay sites are built on One API or a derivative such as New API or sub2api. The same program with a new skin and a new domain is a new site.

Quantity is not diversity. What decides a site's stability floor and price floor is where its quota comes from. Split on that dimension and there are only three kinds.

Kind one: official-quota resale

This kind uses genuine official API keys, whether self-registered, bought at an enterprise rate, or purchased in bulk from a reseller. It does not mix in reverse-engineered traffic.

• Cost structure: the official price times the discount it buys at, usually between 30% and 70% off depending on volume.
• Price floor: it cannot go below 30% of the official price. A site advertising a 0.05 rate while claiming pure official resale can be ruled out on that alone.
• Stability floor: the same as the official channel. When the official endpoint goes down, so does the site, and official rate limits can spill over too.
• Typical risk: quota pulled away by the upstream, leaving the site suddenly unusable, and prices that move with the upstream.

Kind two: mixed-source routing

This kind pools official quota, quota bought from other relay sites, and some reverse-engineered capacity, then routes programmatically by weighting cost, balance and recent success rate. The open-source project metapi is exactly this idea: a relay site for relay sites.

• Cost structure: dynamically weighted, with a natural bias toward sending requests down the cheapest channel.
• Price floor: can be very low, but it passes the cost pressure on to quality.
• Stability floor: depends on the routing policy. A good implementation switches channels invisibly when one fails; a poor one routes you to the worst channel during peak hours.
• Typical risk: the same model name is a different backend at different times. This, rather than deliberate fraud, is the main source of quiet-downgrading complaints.

Kind three: reverse-engineered subscription accounts

This turns web subscription accounts such as Claude Pro or ChatGPT Plus into a standard API. Open-source projects like sub2api do exactly this. In essence it is ride-sharing.

• Cost structure: the subscription fee converted into tokens, so marginal cost is extremely low and the quoted price can sit far below official.
• Price floor: down to 5% to 15% of the official price.
• Stability floor: the worst of the three. Accounts being risk-controlled, banned or rotated shows up directly in your requests.
• Typical risk: first, it violates the upstream terms of service, so bans are routine; second, session stickiness and concurrency ceilings are both bounded by the account pool; third, a single rule change upstream can take the whole service offline at once.

How to tell which kind a site is

You do not have to ask the operator, who may not answer honestly anyway. Look at four observable signs:

Two harder tests. Check whether it supports the Batch API and prompt caching. Reverse subscriptions almost never can, because a subscription account has no such capability. Check how it names model versions. Official resale states the version explicitly, such as claude-sonnet-4-6, while reverse sites usually give only the family name.

What each kind is actually good for

• Critical production paths: consider only the first kind, and treat price as a secondary variable.
• Batch offline work such as data cleaning or content generation: the second kind gives the best value, but you must accept quality variation and sample the results.
• Personal trials, learning, non-critical use: the third kind is enough, provided you never top up a large amount.

The one-line conclusion

The first step in choosing a site is not comparing prices, it is deciding which kind it is. Price is the effect; the source is the cause. Once you know the cause, whether the price is reasonable and where the stability floor sits both follow.

Every site in our directory has its verification status and underlying engine marked on its detail page. Uptime and latency are our own measurements, while rates and minimum top-ups are the site's own claims. Keeping the two apart is what lets you make this judgement yourself.