GetAPIGetAPI.ONE
Choose a model for your task
Filter the current catalog by declared endpoint and task, run a small evaluation with your own examples, and choose from measured quality, latency, and live cost evidence.
Define the decision first
- Write the task, required input/output shape, language, latency tolerance, and failure cost.
- Choose the protocol your application already supports: Responses, Chat Completions, image, video, or another declared endpoint.
- Prepare representative examples and a human-checkable pass condition.
{
"task": "extract support ticket fields",
"endpoint": "openai-response",
"must_pass": ["valid JSON", "required fields", "no invented facts"],
"measure": ["quality", "latency", "live request cost"]
}Build a defensible shortlist
- Open the current catalog and filter by the exact declared endpoint your client will call.
- Read each candidate’s sourced description and live pricing; treat missing facts as unknown.
- Run the same small, representative evaluation set against each candidate when you accept possible billing.
- Record output quality, errors, latency, and observed cost under the same conditions.
- Select a primary model and a fallback only if both pass the required contract.
Use an evaluation checklist
| Dimension | What to record |
|---|---|
| Contract fit | Declared endpoint, required parameters, output parsing, and error handling. |
| Task quality | Pass/fail against examples you reviewed, not a marketing ranking. |
| Latency | Median and tail observations from the same region and request shape. |
| Cost | Live pricing plus actual usage records for the evaluation. |
| Reliability | Errors and retry outcomes during your own test window. |
Verify the expected decision
- Every shortlisted model has the declared endpoint your integration uses.
- The selected model passes all required examples, not merely the easiest prompt.
- The choice document records observation date, test inputs, configuration, and measurable tradeoffs.
Avoid selection mistakes
- Do not infer endpoint support from a familiar model name; use the current catalog declaration.
- Do not compare models with different prompts, regions, output limits, or retry policies.
- Do not turn a vendor benchmark into proof of GetAPI.ONE routing, price, availability, or your task quality.