AI glossary
Effort (reasoning effort)
A setting that tells a reasoning model how hard to think before answering: more effort means better answers on hard problems, but slower and more expensive.
Reasoning models can spend more or fewer tokens thinking. Instead of a fixed budget, many APIs now expose levels such as low, medium and high, and the model adapts how much it reasons. Anthropic, for instance, uses medium as the default on Opus 5.5 and high on Fable 5.1.
Start with the default. Lower effort for simple, high-volume tasks; raise it only where your tests show it helps.
Example: To classify 10,000 emails you use low effort: the task is simple and the model answers quickly and cheaply. To review a complex contract you raise it to high and accept waiting longer.
In practice
- Compare results at several levels on a few real cases before settling on one.
- On simple tasks, more effort is not always better: it can take longer without improving anything.
- You can use a different level on each request depending on what you ask.
Opus 5.5’s new default is covered in the Opus 5.5 and Sonnet 5.5 story.