Claude Code now enables extended thinking for models in the "mythos" family (such as claude-mythos-preview) where it was previously disabled. Two distinct thinking budget paths are updated:
- The always-disabled thinking path (
[false, 0]) now returns[undefined, 2048]for mythos-family models, activating thinking with a 2048-token budget. - The main thinking configuration path now checks for mythos-family models first and applies a dedicated 2048-token budget constant (
Cx4) before falling through to the general thinking settings.
This means users who have claude-mythos-preview configured as their model will now benefit from extended reasoning where Claude thinks through problems more carefully before responding.
Evidence
Model family detection uses ZW_ = ["fruitcake", "macaroon", "mythos"] — search for "mythos" in the source; thinking budget is Cx4 = 2048 (search for "Cx4 = 2048").
Strings lifted out of the shipped bundle, so the claim above can be checked against them.