Vibeleaderboard
← All Intel
Intel / article

The Price of Thinking: Reasoning Effort as a Model-Specific API Contract

Source
arxiv.org
Author
Yeabin Moon
Date

Context

Explicitly requesting high reasoning effort cost more without a detected accuracy gain in a registered test of Sonnet 5. The researcher compared high effort with omitting the effort setting, using 30 AIME 2026 mathematics problems and five calls per problem. The study reports an additional $0.01031 per call under high effort.

The estimated accuracy difference was 1.33 percentage points, with an interval ranging from a 2.67-point loss to a 4.67-point gain. A useful gain therefore remained possible. The comparison concerns those two API configurations on that task; omitting the setting should not be interpreted as switching reasoning off.

Recommended reads
Comments

Checking sign-in…

Loading comments…