---
schema: 1
type: term
title: adaptive thinking
description: >-
  In the channel's model-testing material, **adaptive thinking** is a feature of
  Claude Opus 4.6 in which the model **decides for itself how long to think**
  about a task. It is cited as a suspected cause of the model's higher variance:
  when Opus 4.6 opts not to think much, it can confabulate, contributing to the
  ceiling-vs-consistency tradeoff between Opus and more predictable models like
  Codex 5.3.
canonical: >-
  https://liquidgold.cc/channel/everyinc/explore/term/adaptive-thinking-50be0086baf5
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/adaptive-thinking-50be0086baf5.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# adaptive thinking

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/adaptive-thinking-50be0086baf5

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

In the channel's model-testing material, **adaptive thinking** is a feature of Claude Opus 4.6 in which the model **decides for itself how long to think** about a task. It is cited as a suspected cause of the model's higher variance: when Opus 4.6 opts not to think much, it can confabulate, contributing to the ceiling-vs-consistency tradeoff between Opus and more predictable models like Codex 5.3.

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.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.