Build a requirements shortlist
Start with the model version and required capabilities: context, output limits, tool calling, structured responses and modalities. Confirm provider-specific support instead of assuming that every host exposes the full model feature set. Keep direct and aggregator routes distinct.
Review operational evidence
Source-reported latency, throughput and uptime may use different windows or test conditions. Missing measurements are not zero. The exchange labels provenance and observation times, so you can distinguish published third-party evidence from independent testing.
Validate before committing
Check provider terms, geographic availability, data retention and any service-tier limitations at the source. Test representative prompts with your own budget and acceptance criteria. Paid automated benchmarking remains disabled in the exchange; the directory does not fabricate independent test results.