Model selection
A practical, platform-neutral framework for choosing a model or service for a task.
Working scaffold: Catalog fields, supported modalities, pricing, limits, and availability vary by platform and must be verified before being stated as ONE facts.
Start with the task
Describe the work the application needs to complete, then compare candidates against relevant constraints:
- Input and output: text, image, audio, video, structured data, or a combination.
- Capability: whether the candidate supports the required operation, such as tool use or a constrained response format.
- Context and limits: how much input and output the task needs and what request limits apply.
- Quality and latency: what evaluation criteria and response time are acceptable for the use case.
- Cost: how the service measures and prices the request, including any non-token or per-request charges.
These are evaluation dimensions, not guarantees that every catalog exposes comparable data. Verify each displayed field's meaning, source, and update cadence.
Test before standardizing
Use representative prompts and expected outcomes. Compare quality, latency, failure behavior, and cost on the same workload. Keep evaluation results separate from a model's marketing description.
When available, link users to a live model catalog and the corresponding model metadata reference. The catalog should be the source of truth for current identifiers and capabilities.