Six common tricks in cheap relay services, and how to spot each one

One: the rate is real, the billing is rigged

The trick: the site says "rate 0.3", but billing adds margin to input and output separately, or charges reasoning tokens at a higher rate on their own.

How to spot it: make one call of a known size (say a fixed 100-character prompt asking for a 200-character answer), then work the bill out yourself from the usage field that comes back and compare it with what the site deducted. If they do not match, walk away.

Two: the model is downgraded

The trick: you request claude-opus-5 and get routed to Haiku or an open-source model. Nothing looks wrong, because the API response format is identical.

How to spot it:

• Set an A/B test with questions that need deep reasoning — same question against the vendor and against the relay service, compare quality
• Compare output token counts: the same model on the same prompt should produce a similar count. A large gap means a different model
• Watch response speed: an Opus-class model cannot return long text in 300ms every time
• Check whether the price is below a plausible floor (a flagship Claude model at rate 0.1, for instance)

Three: the credit is shrunken

The trick: you pay ¥100 and are credited 1,000,000 units, but converting at the vendor list price shows you can only actually spend about 60% of what was advertised.

How to spot it: immediately after topping up, make one metered call and reverse the actual units consumed from the list price. This takes under a minute and filters out a fair number of services.

Four: rate limits as a disguised price rise

The trick: the cheap tier is capped at one or two concurrent requests, throws 429 the moment you push, and then pre-sales tells you to "upgrade to the VIP route for unlimited speed".

How to spot it: top up a small amount and run a ten-concurrent script, then count the success rate. If the cheap tier is effectively unusable, the "low price" is a lead magnet and the real cost is the VIP price.

Five: big top-up bonuses that trap your money

The trick: pay 1,000 get 500 free; pay 10,000 get 5,000 free. A bonus ratio that high exists to keep your money on their books.

How to spot it: normal businesses offer roughly 5%–15%. A bonus above 50% means they need new users' money to cover existing users' credit — that is the signature of a Ponzi structure. However tempting the bonus, never put in more than you can afford to lose.

Six: they vanish

The trick: no long-term operation. Collect a round of top-ups, close the site. Some give warning by announcing a "system upgrade" first.

How to spot it (this lowers the odds, it does not remove them):

• Look at the uptime history. Our uptime leaderboard flags services below 90% over a sustained period
• Look at how long they have operated and how active their public community is. Services with no traceable operating entity are the highest risk
• Look at whether they can issue invoices or accept corporate transfer. A service willing to go through formal processes has more to lose by disappearing
• The most effective measure: only ever top up one or two months' worth.

A general self-protection routine

1. Shortlist two or three candidates, all filtered by uptime on our rankings page
2. Top up the minimum at each
3. Run three tests: a basic call, a tool call, ten concurrent requests
4. Reconcile one bill
5. Only then consider adding volume — and always top up in batches

Where we stand

We do not move a service up the table because it paid us. Ordering is decided by measured data alone. If you run into any of the behaviour above at a listed service, tell us through submit a correction. Once verified, we remove the service and keep the record.