AI API Relay Review Methodology
Read how AI Relay Check evaluates uptime, actual cost, model identity, task performance, support, and commercial independence.
Evidence levels
Paid model requests, website probes, provider-published material, third-party records, and unverified submissions use different labels. Only timestamped model requests count as API verification.
Failure threshold
One model failure remains a sample. Verification is removed only after every configured model in the same provider family fails at least four consecutive times.
Pricing rules
Input and output rates remain separate, as do pay-as-you-go prices, full-use subscription equivalents, and observed wallet charges. Different routes and client restrictions do not share one price rank.
Commercial independence
Affiliate commissions, advertising, and provider test credit do not change samples, score weights, or ranking order.
Evidence layers and decision boundaries
| Evidence | How collected | Stored fields | Does not prove |
|---|---|---|---|
| Website reachability | Scheduled HTTPS request | HTTP result, final hostname, timestamp | Does not prove API availability |
| Model API verification | Paid request with a dedicated test key | Model ID, protocol, success/failure, latency, tokens | Does not prove model identity |
| Published pricing | Provider-owned page or endpoint | Input/output rate, group, capture date, source URL | Does not equal observed billing |
| Billing reconciliation | Fixed request plus balance before/after | Tokens, cache fields, expected and observed charge | Valid only for the tested route |
| Model identity | Repeated capability and consistency probes | Cross-run behavior and control-task evidence | Never inferred from the response model name alone |
Failed paid requests remain in the sample. Price, uptime, API success, billing, and model identity are never substituted for one another. Affiliate links and test credit do not change score weights.