Claude Code Effort Levels, and Where Ultracode Fits
Effort controls how hard Claude thinks; Ultracode controls how many agents it can hire. The /effort menu offers a ladder from low to xhigh plus a separate Ultracode toggle. Since Claude Code v2.1.284 they are two independent dials — you can run deep reasoning without orchestration, or broad orchestration at modest effort.
The effort ladder
Open it with /effort — you get a slider:
| Level | Think of it as | Typical use |
|---|---|---|
low |
Fast and cheap | Renames, small edits, quick questions |
medium |
Balanced default | Everyday feature work. Default on Opus 5.5, the current default Opus |
high |
Careful | Tricky bugs, multi-step refactors |
xhigh |
Maximum reasoning | Architecture calls, gnarly concurrency bugs — the level tied to Ultracode workflows |
Two model-dependent details worth knowing:
- xhigh isn't universal. The toggle only appears for xhigh-capable models — current-generation ones like Opus 5.5 and Sonnet 5.5. On Opus 4.6 or Sonnet 4.6 you won't see it.
- Defaults differ by model. Opus 5.5 ships with
mediumas its default effort.
The second dial: Ultracode
The same /effort menu has an Ultracode toggle (flip it with Tab). That toggle turns on automatic workflow orchestration — Claude writes and runs multi-agent scripts whenever a task warrants it. Full setup walkthrough in how to enable Ultracode.
The mental model that keeps them straight:
Effort scales depth in one context. Orchestration scales breadth across many. xhigh makes one line of thinking better; Ultracode adds more lines of thinking in parallel.
The launch flag exception
claude --effort ultracode is the one place the two dials fuse: it starts the session with Ultracode on and effort pinned to xhigh. Inside the session, /effort ultracode changes only the orchestration toggle.
Organization caps
Team and Enterprise admins can set an organization-level effort cap. The cap lowers the effort an Ultracode session runs at but does not disable Ultracode itself.
How the dials affect cost
- Raising effort raises tokens per turn — linear and predictable
- Turning on Ultracode multiplies agents per task — that's the big lever. A single workflow can burn hundreds of thousands of tokens across dozens of subagents
- Low effort does not make fan-outs cheap. Agent count drives workflow cost far more than per-agent effort does
- Keep the blast radius sane with the size guideline:
/config workflowSizeGuideline=small(<5 agents),medium(<10),large(<50), orunrestricted. Default issmallon Pro plans
Benchmark before you commit
If you're deciding whether Ultracode earns its keep on your workload, run an honest 2×2: effort × Ultracode, fresh session per arm, fixed size guideline, and report token counts per arm. Pick the cheapest quadrant that passes your quality bar — on many tasks that's high effort with Ultracode off, and orchestration reserved for genuinely broad jobs like audits and migrations.