Reasoning effort values

Reasoning effort is set with --reasoning-effort <value>, /reasoning-effort (alias /thinking), or [generation] reasoning_effort in ~/.magic/config.toml.

The conceptual usage guide lives at Reasoning effort. This page is just the table.

Levels

Value	Search aliases	Hint
none	off, disabled, disable, anthropic no thinking	Disable reasoning. Anthropic mapping: no thinking.
minimal	min, fast, anthropic no thinking	Lightest OpenAI reasoning. Anthropic mapping: no thinking.
low	—	Lower latency, lower token use.
medium	balanced, normal	Balanced reasoning depth.
high	deep, strong	More thorough reasoning.
xhigh	max, maximum, strongest, deepest, anthropic max	Strongest OpenAI level. Anthropic mapping: native xhigh on Claude Opus 4.7, else max when supported, else high.

Accepted at the CLI

--reasoning-effort accepts the canonical values plus max (which is normalized to xhigh):

none, minimal, low, medium, high, xhigh, max

Anything else is rejected with an error.

Accepted in /reasoning-effort

The slash-command picker fuzzy-matches against canonical values, search aliases, and the hint text. So all of these work:

/reasoning-effort high
/reasoning-effort max          # surfaces xhigh
/reasoning-effort balanced     # surfaces medium
/reasoning-effort fast         # surfaces minimal
/reasoning-effort default      # clears the override

Stored in config

~/.magic/config.toml:

[generation]
reasoning_effort = "high"

Magic Coder normalizes the stored value to one of the canonical names. Storing max is allowed and read back as max; the runtime treats it as xhigh.

Cross-provider mapping

• OpenAI reasoning models — values map directly: none, minimal, low, medium, high, xhigh.
• Anthropic models with extended thinking — none and minimal disable thinking; low/medium/high scale the thinking intensity; xhigh stays native on Claude Opus 4.7, otherwise it maps to max when supported, else high.
• Other providers — the closest equivalent the model supports.

You don’t need to remember the mapping per provider — pick the level that matches the difficulty of the task and Magic Coder handles the translation.