Models, effort & ultracode

Claude Code has four dials that change how it works. Picking them well saves time and money:

• Model: which Claude you talk to.
• Effort: how hard it thinks before answering.
• Ultracode: whether it splits big tasks across many helper agents.
• Fast mode: the same Opus quality, answering faster, at a higher price.

Facts below come from the official docs (model config, workflows, fast mode, ultrareview), checked on 2026-10-21. Settings change over time, so re-check the docs.

Model (switch with /model):

Model	Use it for
haiku	Simple, quick tasks
sonnet	Everyday coding
opus	Complex reasoning
opusplan	Opus plans, Sonnet does the work
fable	Work that spans more than one sitting: root-cause investigations, outages, architecture. Not the default; you have to pick it.

Effort (set with /effort or /effort <level>):

Level	Use it for
low	Quick edits you'll review yourself
medium	Routine or cost-sensitive work
high	Bug fixes where getting it right matters
xhigh	Deeper reasoning, at more cost
max	Hard problems the agent works on alone. Watch for overthinking.

The default is high, or medium on the 5.5 models.

Ultracode is not an effort level. It's a setting that makes Claude plan a workflow for big tasks: a script that runs many helper agents. It runs at whatever effort you've set.

• Turn it on for the session: /effort ultracode. Starting with claude --effort ultracode also sets effort to xhigh.
• For one task only: put the word ultracode in your prompt.
• Cost: more tokens and more time per task, so you hit plan limits sooner.

/code-review ultra is a separate feature: a deep, multi-agent bug review that runs in the cloud and takes roughly 15–25 minutes. A few free runs, then usage credits.

Fast mode (/fast) works on Opus only. It's up to 2.5× faster with the same quality, at a higher price, billed from usage credits. Turn it on at the start of a session, because switching mid-conversation re-charges the whole conversation once. It's good for live back-and-forth and wasteful for long unattended work.

/effort          (opened it, then cancelled: I was just looking)
/code-review max (the built-in bug hunt at max effort, in the background)

The whole site was built on Opus 5.5 at the session's default effort.

How I'd pick, after building this site (my rule of thumb, not official guidance):

Situation	Model	Effort	Ultracode
Grilling, planning, spec	opus	high	off
Small, clear issue (scaffold, CSS)	sonnet	medium	off
Test-first feature with real logic	opus	high	off
Mystery bug across many files	fable or opus	xhigh	maybe
Final review before shipping	opus	max	off
Huge refactor or broad research	opus	xhigh	on

The lesson: higher isn't always better. More effort and ultracode cost more and take longer. Match the dial to how hard the problem is, and how costly a mistake would be.