Models, effort and fast mode

You know the basics from the lesson on the model and effort: the model is the ceiling on capability, effort is how much work goes into an answer. Here is what comes up once you configure Claude Code seriously: which layers set the choice, what overrides what, and where extra spending hides.

Three more aliases

Add these to sonnet, opus, haiku and opusplan:

Three important details:

• An alias moves. Over time it starts pointing to a newer model. If you need one specific version, write the full model name from Anthropic’s models overview.
• default is not a model. /model default clears your choice and returns the default model for your account.
• The provider decides. On Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, the same alias may lead to an older model. And depending on your plan, Fable may draw on usage credits.

Effort: what overrides what

Effort is set in several places, and each lasts a different length of time:

Above all of them sits the maxEffortLevel ceiling: not even the environment variable breaks through it. /effort auto clears the level saved for the current model — then effortLevel from settings applies if it fits, otherwise the model’s own level. /effort status shows which level is in force right now.

Fine points that catch experienced users:

• max is not saved in files. effortLevel and modelSettings do not accept it. Through /effort it lasts one session; permanently, only through CLAUDE_CODE_EFFORT_LEVEL.
• A top-level effortLevel is outdated. Set at the top level of user settings, it does not apply to the newest models. For a specific model, use modelSettings.
• Models have different levels. Ask for a level a model does not have, and you get the nearest available one below it. And high on one model is not the same amount of reasoning as high on another.
• ultrathink is for one turn. The word in a request asks for deeper thinking on that turn; the level does not change. “Think really hard” is ordinary text.
• Ultracode is not an effort level but a separate setting: with it, Claude plans a dynamic process for every serious task at whatever level the session is running.

An example of targeted tuning — a skill where checking matters gets its own level, independent of the session:

---
name: five-whys
description: Finds the cause of a failure from the facts.
effort: high
---

While the skill is active, it runs at high. The CLAUDE_CODE_EFFORT_LEVEL variable and the ceiling still take precedence. And note: Haiku is not supported in auto mode, so a skill with model: haiku stays on the session’s model there.

Fast mode

/fast is the same Opus with an API configuration that answers faster but costs more per token. The quality is the same. The documentation calls the mode a research preview: features and prices may change.

What to know before you turn it on:

• Opus only. The mode works on supported Opus models; if you were on another model, /fast switches you to Opus. After turning it off you stay on Opus — switch back with /model.
• Always usage credits. On a subscription, fast mode is paid only from extra usage credits — even if your plan’s limit is not used up. Without credits enabled, /fast answers Fast mode requires usage credits.
• Turn it on at the start. If you turn it on mid-conversation, the whole conversation up to that point is billed once at fast mode’s uncached price.
• Watch for the icon. When the mode is on, ↯ appears next to the input.
• Not on every provider. It is not available on Bedrock, Google Cloud, Microsoft Foundry or Claude Platform on AWS.

Where it pays off: live debugging and fast iteration, when waiting costs more than tokens. Where it does not: long autonomous runs, CI and anything where cost matters. On Pro and Max, spending shows in the Usage credits section on claude.ai; on Team and Enterprise, through /usage.

And about numbers in general: total_cost_usd in the JSON from claude -p is a client-side estimate, not your bill.