Choosing model, effort and fast mode

Three dials, not one

People talk about "using the strong model" as if it were a single choice. In Claude Code it is three, and they are independent.

Model decides who is thinking. Effort decides how hard that model is asked to think on a turn. Fast mode decides how quickly the same model answers, and at what price. Setting each one for the right reason is worth more than always turning all of them up.

The model

You pick a model with an alias, and aliases move forward over time. The list is default, best, fable, sonnet, opus, haiku, plus sonnet[1m] and opus[1m] for the one-million-token window, and opusplan. If you need a model that never changes under you, pin a full model ID instead of an alias.

opusplan uses opus while planning and sonnet while executing. It is a good fit for the work, and you should know the cost: every plan-mode toggle becomes a model switch, and a model switch resets the cache.

The order in which a choice is applied runs from your most recent decision to the most general one: /model inside the session, --model at startup, the ANTHROPIC_MODEL variable, the model key in settings, and ANTHROPIC_DEFAULT_MODEL.

Effort and thinking

The effort levels are low, medium, high, xhigh and max, and the command line also accepts ultracode. You set it with /effort or --effort.

The guide's advice reads like a ladder:

• medium for everyday work.
• high or xhigh for subtle bugs and schema changes.
• low for quick back and forth.
• max only for hard, unattended problems, because it tends to overthink.
• ultrathink in a single prompt for one tricky question.

The guide is explicit that phrases like "think step by step" are plain text, not keywords. Effort is the real control. The one wording that still helps is "be ambitious", because it signals scope rather than thinking depth.

On those models, changing effort keeps the cache. On most others it does not.

Fast mode

Fast mode makes Opus up to 2.5 times faster, at a higher price per token. The documented figures are 8 and 40 USD per million input and output tokens on Opus 5.5. It is not a different model. It is the same one, served faster, which is exactly why it has a separate cache.

On subscription plans it is billed through usage credits, not plan limits. A small lightning icon shows while it is on. /fast toggles it, and the setting fastMode: true persists it.

Timing matters here as much as in the cache lesson. If you turn it on mid-turn, the running turn finishes at normal speed. If you turn it on mid-session, the first enable reads the whole conversation uncached at fast rates. Turn it on at the start or not at all.

A way to decide

Choose at the start, and choose by the task.

The last line is the point of this whole course. A bigger dial does not fix a window full of the wrong things. When an answer is poor, look at the context first.

If you want a second opinion, /advisor or --advisor <model> lets a stronger model advise the session.