Start from the workload
Write down the calls your product actually makes: model IDs, streaming requirements, image or video task polling, tool calls, structured outputs, concurrency, and timeout expectations. A provider can support an API shape in general while still missing one detail your application depends on.
Create a small fixture set from sanitized production-like requests. Keep the input, parameters, expected response fields, and acceptance criteria together so every candidate sees the same test.
Measure four independent dimensions
Score evidence, not promises
Record the test window, runner region, request count, concurrency, model identifier, retry policy, and raw result location. Separate measured values from vendor documentation and clearly date both. A single prompt is a workflow check, not a universal model ranking.
Use the open-source gateway evaluation kit and provider benchmark as starting points.
Decide with failure drills
Before launch, deliberately test invalid credentials, unknown models, oversized inputs, timeouts, 429 responses, and provider-side failures. Confirm that your application classifies these cases, retries only when safe, and preserves enough context to investigate the request.
Use supported models through one APIMART account.
Confirm current model availability, pricing, and limits before routing production traffic.
Create an APIMART account