A test table can narrow your choices, but every number has a scope. Start with the time window, sample size and testing method.
Uptime
Uptime is the proportion of successful test requests. It does not promise zero outages or guarantee your own calls will work. Review models and time periods separately.
Latency
Response time depends on the model, output length and network path. Comparable tests need the same prompt and output limit. For streaming APIs, both time to first token and total duration matter.
Sample size and freshness
Ten tests with 100% success cannot be treated like thousands of tests with 99% success. Old data should be marked stale and removed from active recommendations.
Model consistency
Response characteristics can signal unusual routing, but one answer cannot prove the underlying model's identity. Treat this as a risk signal, not definitive proof.