Model selection can no longer be reduced to “use the strongest one”. You must choose the model, reasoning effort, and task class together. Otherwise the system either overpays for routine work or gets shallow output where verification is needed.
The common mistake is not the model name. It is the effort setting.
Too much effort on simple tasks burns time and budget. Too little effort on migrations, reviews, or bug investigations produces work without proof: the model may skip files, leave hypotheses unchecked, or stop before a reproducible result.
Practical rule
First define the task type and success criterion. Then choose the cheapest mode that reliably passes the check.
For an editorial pipeline:
- low effort for first-pass email sorting;
- medium for source summaries and short verification;
- high for complex synthesis with several sources and counterarguments;
- max only where error costs more than computation.
The same logic is visible in coding agents. Quality is determined by the whole chain: effort, tools, tests, review, and human confirmation.
New Runtime Read
Model match is not choosing the “best model”. It is choosing the minimum mode that reliably passes the required check.
For production systems, routing should be part of the design, not an operator habit. Cheap reversible tasks do not need the most expensive mode. Risky tasks can turn reasoning-depth savings into rework and hidden errors.