Compare the system, not one adjective
| Question | Smaller model may suit | Larger model may suit |
|---|---|---|
| Task structure | Narrow, repeated, well specified | Ambiguous, varied or multi-step |
| Inputs | Stable formats and vocabulary | Messy context or nuanced instructions |
| Operations | High volume or strict latency | Lower volume where difficult cases dominate |
| Risk | Easy to validate or safely reject | Complex judgment, still with appropriate verification |
These are hypotheses, not universal laws. Architecture, training, tooling and task fit can make a smaller model outperform a larger one on a bounded evaluation.
Use a capability portfolio
Instead of declaring one organisation-wide winner, map task classes to the simplest reliable method. Deterministic code can enforce rules, a small model can handle routine language variation, and a more capable model can receive ambiguous cases. Human escalation remains part of the design where consequences require judgment.
Keep the routing explanation simple enough to observe. If no one can tell why a case escalated or failed, savings will be difficult to defend.
Evidence & context: Microsoft Learn
Run a blind comparison
- Create representative cases and an acceptance rubric.
- Remove model names from outputs during review where practical.
- Record pass rate, severe failures, latency and full task cost.
- Inspect results by task category, not only as one average.
- Choose a fallback and a date to re-evaluate.
Sources & further reading
- Models
OpenAI Developers. Official model-selection documentation, checked 13 September 2026. Product names, capabilities and prices can change; the collection uses the durable principle of matching capability to a task rather than prescribing a current model.
- Evaluate a model router
Microsoft Learn. Official guidance for evaluating routing across representative workloads using quality, cost, latency and policy criteria. It is not evidence that routing always improves results.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗