Where the term comes from In the channel's coverage of the Opus 4.6 vs. GPT-5.3 Codex head-to-head, adaptive thinking names a feature of Claude Opus 4.6: the model decides for itself how long to think on a given task, rather than following a fixed reasoning budget. Why it matters in practice The team treats adaptive thinking as a suspected factor in Opus 4.6's behavior profile — part of why it pairs the highest ceiling of any coding model they have used with the highest variance. When the model decides not to think much, it sometimes goes off the rails and confabulates. That makes adaptive thinking a concrete piece of the broader ceiling-vs-consistency tradeoff the chapter explores: Opus can reach peaks no other model reached, but its self-directed reasoning depth makes it less predictable than Codex 5.3, which is consistently very good and safer to leave unsupervised. - Adaptive thinking = Opus 4.6's ability to choose its own reasoning length per task.- It is suspected (not confirmed) as a source of Opus 4.6's confabulations when it under-thinks.- It feeds the chapter's core lesson: pick models by variance tolerance, not just peak capability.