ARC-AGI-3 API clarification: 'reasoning' field logs model output, not private CoTs

GregKamradt · x · 2026-07-30

Greg Kamradt clarified a common confusion regarding the ARC-AGI-3 evaluation API. The optional reasoning field is used to log arbitrary outputs and metadata associated with a model's actions for analysis in the replay viewer. It does not capture the model's internal or private Chain-of-Thought (CoT). The author noted plans to rename this field to something like 'action metadata' to prevent future misunderstandings.

Related event: ARC-AGI-3 Officially Open-Sources Benchmarking Codebase(4 posts)→

Original post →

More from coding & agent

coding & agent channel →