Learning outcomes
- Compare model classes
- Plan version changes
- Define selection criteria
Mental model
A model name is a versioned dependency with capabilities, modalities, limits, prices, latency, safety behavior, and a lifecycle, not a permanent intelligence tier.
Theory
Selection should start with task-level acceptance tests and operational constraints. Larger or reasoning-oriented models may improve difficult cases but increase latency or cost. Smaller models can handle routing, classification, and extraction. Pin versions where supported, record model identity, maintain an upgrade suite, and plan for deprecation.
Alternatives and trade-offs
A single-model architecture is simple; routing can improve economics and resilience but adds evaluation and observability requirements. Hosted and self-hosted models shift control and operational burden.
Failure modes and misconceptions
Do not select from benchmark reputation alone, assume aliases are immutable, or upgrade without replaying representative and adversarial tests.
Knowledge check
What evidence should justify moving a workflow to a smaller model?
Decision scenario
Route high-volume classification to a smaller validated model and escalate uncertain cases, while tracking disagreement, latency, cost, and drift.
Relationships
Primary sources
- OpenAI Models - OpenAI, verified 2026-07-16