The labels are not standardized across providers
One API exposes an explicit reasoning-effort field. Another may encode a tier such as Max in the model ID while returning a default effort value. Those two representations should not be treated as the same control.
This is why ModelDial keeps the raw model identity, normalized display name, effort, and route in the configuration record instead of flattening them into one brand row.
Test the effort curve within one configuration family
Hold the model, route, and comparison standard fixed. Compare Low with Medium, then Medium with High. If the route changes at the same time, you cannot tell whether the result came from reasoning effort or transport behavior.
- Check individual task results as well as the overall score; an unchanged total can hide differences.
- Compare successful elapsed time, not only time to first token.
- Without repeated comparable observations, a small gap does not establish a stable lead.
When higher effort is worth the extra time
A higher setting may take longer without improving the relevant question, and ordinary model variance can still reorder adjacent efforts in one batch. Higher effort is therefore a hypothesis to test, not an automatic upgrade.
For a main agent, keep the lowest effort that consistently clears the quality floor. For bounded worker tasks, measure the task-specific floor separately instead of inheriting the main agent setting.
Record the exact setting you can reproduce
Save the provider, raw model ID, effort, service tier, route type, and evaluation time. A label such as High is not enough to reproduce a result after an endpoint or provider configuration changes.
What changes at the next effort level?
See gpt-6-astra's scores, reference costs and evaluation times at neighboring tested efforts. Switch pairs to see where more effort helps.