TL;DR
The short version
- Red flags: unnameable key source, shared quotas, below-cost prices, no usage ledger, anonymous payment only, model-swap suspicion, zero operating history.
- Failure modes: upstream ban (everyone dies at once), silent model downgrade (降智), exit scam, quota starvation.
- Green flags: names its upstream, per-account keys, sustainable pricing, itemized billing, contactable support — the standard NovAI publishes for itself.
What a Relay Actually Is
A relay — 中转站 in the Chinese developer community — fronts a model provider's API with its own endpoint and resells access. Not every reseller is bad: a legitimate gateway holds official commercial access, issues you your own key, and bills you transparently. The dangerous ones pool a handful of upstream keys across thousands of customers and undercut real cost to grow fast. The math never works forever — and when it stops working, your production traffic is collateral damage.
The 7 Red Flags
1. They can't (or won't) name the source of their access.
"How exactly do you obtain your API capacity?" If the answer is evasive — "our channels", "partnerships" with no name — assume the access is borrowed, not contracted.
2. Pricing far below the upstream's own cost.
If the model's official rate is $1.00 per million output tokens and the relay sells $0.30 "unlimited", someone is subsidizing you. It's either shared quota (flag 3) or a countdown to an exit scam.
3. Shared keys and shared quotas.
Random 429 storms, throughput that varies by hour, "please limit your concurrency" messages — all signs your key is a timeshare. When the upstream bans that key, every customer on it dies simultaneously.
4. No itemized usage billing.
Balance goes down, but you can't see per-request costs, tokens, or timestamps? Then you can't detect overcharging, and the provider can't prove it isn't. Auditable billing is table stakes.
5. Quality drifts while prices stay flat.
The classic 降智 ("dumbed-down") pattern: your flagship-model calls quietly route to a cheaper smaller model. The tell is refusal to provide request-level logs or upstream response headers.
6. Anonymous payment rails only, no refund policy.
Crypto-only checkout with no invoicing, no terms, no refund path is a structural exit-scam enabler. Legitimate providers offer mainstream rails (card, PayPal) alongside crypto.
7. Zero operating history.
Domain registered last month, no changelog, no status page, support via a single Telegram account. Longevity is cheap to fake for a month and expensive to fake for years — check the age.
The Four Ways Relays Fail
| Failure mode | What happens | Your damage |
|---|
| Upstream ban | Pooled key violates the provider's terms; access revoked | Hard outage, no timeline, no recourse |
| Silent downgrade | Flagship calls routed to cheaper models | Product quality degrades invisibly |
| Quota starvation | Too many customers on too few keys | Latency spikes, random failures at your peak hours |
| Exit scam | Deposits collected, site vanishes | Balance gone, keys dead |
Your Due-Diligence Checklist (works for any provider)
- Key source: can they name the upstream and the nature of the access (official channel vs pooled keys)?
- Dedicated quota: is your key yours alone, with your own rate limits?
- Sustainable pricing: at or above realistic upstream cost — deep discounts need a credible explanation.
- Auditable billing: per-request logs you can reconcile against your balance.
- History & contact: operating track record, real support channel, terms and refund policy.
For the record, how NovAI answers these: official direct channels from the providers (ByteDance, Alibaba, Zhipu, Tencent), per-account keys with no shared quotas, provider list pricing with no markup, itemized auditable billing, and USDT (TRC20) alongside standard rails. The same questions are on our FAQ — ask any vendor, including us.
FAQ
What is an AI API relay?
An unofficial reseller that fronts a provider's API. The risky kind pools shared upstream keys across customers, undercuts real cost, and keeps no auditable billing.
Why are shared API keys dangerous?
Quota is enforced per key: you compete with strangers for throughput, and one upstream ban takes everyone down at once.
Can a relay secretly serve a weaker model?
Yes — the 降智 pattern. Watch for quality drift plus a refusal to show request-level logs.
How do I verify a provider is legitimate?
Named upstream source, dedicated quota, sustainable prices, itemized billing, real operating history. Five checks, five minutes.