Is an AI API Relay Reliable? Multiplier, Exit Risk, Fakes
Price alone is not enough. Learn to read multipliers, spot exit-risk signals, see why intermittent fakes are riskier, and follow a verification checklist.
Discussions about choosing a relay often stop at price comparison. But price is the easiest variable to see, and the one that is least safe to interpret on its own. This article covers what you should actually look at.
How to Read the Multiplier
"Multiplier" is a cost coefficient relative to official pricing. A 0.7× multiplier means you pay 70% of the official price.
A multiplier by itself is not evidence of wrongdoing, but it is a number that needs explaining. Below 0.5 means the provider gets quota at a cost far below the official API price, which may come from:
- Reverse-proxying the internal interfaces of other AI products (low cost, but they come with hidden system prompts)
- Subscription account pools (low cost, but with rate and concurrency limits, and the contract risk sits with the seller)
- Fake models (lowest cost, because you are not actually buying the model at all)
The first two are business model choices; the third is fraud. A low price is not the problem. A low price with no explainable source is the problem.
Signals That Predict Exit Risk
These are not verdicts. They are signals that warrant one more question:
| Signal | Why it matters |
|---|---|
| Accepts only bank transfers and cannot issue invoices | Lacks a traceable business entity; no path for recourse if something goes wrong |
| Recruits customers only on social platforms, with no independent site or terms of service | Cost to enter and leave is extremely low; shutting down has no sunk cost |
| Large prepayments with deep discounts, no minimum spend, no refund terms | The larger the prepayment, the stronger the incentive to disappear |
| Terms of service do not state upstream sources or SLA | If something goes wrong, "we changed the model" is not a breach |
| Short operating history, no long-term third-party monitoring records | No verifiable performance history |
Here is a blunt recommendation: do not make a large purchase from a new site at once. No matter how big the discount, whether you can get service depends on whether the provider is still open.
Fakes Are More Dangerous Than Total Failure
An endpoint that does not work at all gets dropped the same day. The real losses come from intermittent problems:
- We tested an endpoint with a pass rate of about 10% and consecutive 503 errors. This kind of endpoint makes you think your retry logic is not robust enough, so you add retries and keep using it
- We have also seen an endpoint whose headers were sent normally and whose TTFT was just 182ms, but whose body was completely empty. If monitoring only watches response time, this kind of endpoint looks fine in every way
Intermittent success = high risk. It does not trigger your alarm; it slowly contaminates your output quality and your bill.
An Actionable Verification Checklist
Run this before going into production, not just on the day you buy:
- Model authenticity — compare behavioral fingerprints rather than trusting self-claims. Sibling models within the same family must be distinguishable (Opus vs Haiku is the most common downgrade direction)
- Token counting — use prompts with known token counts to check the returned
usage. Slight inflation (5–15%) is invisible to the eye - Hidden system prompt — measure the gap between the claimed input and what is actually billed
- Streaming format — whether SSE is compliant, and whether there are cases of zero chunks while reporting normal
- Long context — whether the claimed context length is actually usable, or silently truncated
- Stability — pass rate under sustained requests, not a single-success rate
- Retesting — an upstream may provide a legitimate channel early in a partnership and switch to a low-cost channel later. Testing only on the day you buy is no test at all
Item 7 is the most often overlooked, and the most likely to cost you.
Use a Tool, Not Manual Testing
We have automated all seven items above. Relay testing accepts any OpenAI-compatible endpoint and, after running, outputs a 0–100 score with item-by-item verdicts. The decision tree is public, including a conservative branch that abstains when evidence is insufficient. A detection tool that never says "not sure" is not trustworthy in itself.
To disclose a conflict of interest: BazaarLink also provides AI API services, so we are participants in this market. This is why we make our verdict logic and benchmark data public, so you can check our methodology rather than having to trust our conclusions.
Further reading: Three methods to detect Claude capability downgrade, Does connecting Claude Code to a relay cause a capability downgrade?, Token padding detection.
FAQ
Is a multiplier below 0.5 always a problem?
Not necessarily, but it needs an explanation. The multiplier is a cost coefficient relative to official pricing. Below 0.5 means the provider's cost to obtain quota is far below the official API price, which may come from reverse-proxying another AI product's interface, a subscription account pool, or fake models. A low price is not evidence of wrongdoing by itself, but an endpoint that sells below cost without a clear source deserves testing before you invest.
Why is intermittent success more dangerous than total failure?
Because you will not notice it. An endpoint that does not work at all gets dropped the same day. An endpoint with about a 10% pass rate makes you think your retry logic is at fault, so you keep using it for months. In our testing, an endpoint with an approximately 11% pass rate and consecutive 503 errors did far more damage than one that simply went down.
TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API