NeuralCodeDOCS
Configuration

Models & thinking

Choose which Claude model your agents use and how hard they think before answering. Both are set in the composer and in Settings → Models.

Models#

Models are grouped by family — Opus (most capable), Sonnet (balanced speed and quality), and Haiku (fastest). Pick the newest in a family, or an older version for reproducibility.

FamilyVersionsBest for
Opus4.8, 4.7, 4.6, 4.5, 4.1The hardest tasks — deep refactors, tricky bugs, ambiguous specs.
Sonnet4.6, 4.5, 4.0Everyday work — a strong balance of speed, quality and cost.
Haiku4.5, 3.5Fast, cheap, routine edits and quick iterations.

The selected model applies to new runs. You can switch models per message right from the composer.

Extended thinking#

Extended thinking gives the model more room to reason before it answers — deeper, but slower and slightly more expensive. Choose a budget:

LevelWhen to use
OffRoutine work where speed matters most (default).
HighNon-trivial changes that benefit from planning.
XHighComplex, multi-file work or subtle logic.
MaxThe gnarliest problems — maximum reasoning depth.
Cost control
Lowering thinking and using a cheaper model for routine tasks are two of the built-in tips on the Usage screen — both can be applied in a click.