Last updated: 2026-04-19

/effort xhigh vs high vs max — Opus 4.7 Effort Levels Explained

Claude Code 2.1.111 (April 2026) shipped xhigh — a new effort level that sits between high and max. The release also added an interactive slider, so /effort with no arguments now opens an arrow-key picker instead of requiring you to remember the exact name. Here's what each level actually does, when to use them, and how the cost breaks down.

The Five Levels

LevelWhat it's forRelative costTypical latency
lowTrivial questions, formatting, lookups. The fastest option.1× (baseline)< 2s
mediumRoutine coding, refactors, doc edits. Good chat default.~1.3×2–5s
highComplex coding, multi-file changes. The right default for most engineering work.~2×5–15s
xhigh NewHard debugging, architecture design, security review. The new sweet spot when high isn't enough.~3–4×15–40s
maxGenuinely novel problems, research-grade reasoning. Rarely necessary.~6–8×40s–2min

Cost and latency are approximate; exact numbers vary by prompt complexity and model. Other models fall back to high when xhigh is requested — only Opus 4.7 honors the new tier directly.

How to switch

Three ways:

  1. Slash command: /effort xhigh — direct switch.
  2. Interactive slider: /effort with no arguments. Arrow keys to move between levels, Enter to confirm. New in 2.1.111.
  3. CLI flag: claude --effort xhigh — set for the whole session at launch.

The setting persists for the session. Use /effort again to change mid-conversation; cost from prior turns isn't affected.

When to use each

Default to high

For day-to-day engineering work, high is the right baseline. It produces multi-file refactors, working test suites, and reasonable architectural decisions without burning through your monthly quota. Most people who feel like Claude Code is "too expensive" are running everything on max when high would have produced equivalent results.

Drop to medium for chat

If you're asking Claude to explain something, summarize a file, or do a one-line fix, medium is plenty. The marginal quality gain at high for these tasks is small and you'll feel the latency.

Bump to xhigh when stuck

The clearest signal: you ran high, it produced something that looked plausible but didn't work, and on retry it produced the same kind of plausible-but-wrong answer. That's the failure mode xhigh targets — problems where the model needs to think longer to actually reason through the constraints. Hard concurrency bugs, design choices with non-obvious trade-offs, security advisories that require correlating multiple files.

max is rarely the answer

It exists for genuinely hard problems — novel architecture decisions, research-grade analysis. If xhigh didn't solve it, the next move is usually to break the problem into smaller pieces, not throw more compute at the same prompt. max is also gated to Max tier subscribers (and now works with Auto mode without the old --enable-auto-mode flag).

The Auto mode interaction

Claude Code 2.1.111 also shipped Auto mode for Max subscribers when using Opus 4.7. Auto picks the effort level for you based on prompt complexity — it's a different lever than /effort and can override your manual setting. If you've set /effort xhigh but Auto decides medium is enough for the next prompt, Auto wins. Disable Auto if you want to enforce a fixed effort.

Practical cost rules of thumb

  • Routines should pin effort. Scheduled jobs running unattended should set effort explicitly (usually medium or high) — Auto mode's variability can blow your budget on a single off-day.
  • Pair Haiku with low/medium for batch work. If you're processing 100 items, use Haiku at medium not Opus 4.7 at xhigh. The cost difference is 50× for marginal quality gain.
  • Don't use xhigh for code review. Use /ultrareview instead — it's the new multi-agent review command from 2.1.111 and it's cheaper than running xhigh against the same diff.

Related

← Back to the Claude Cowork FAQ

📬 Weekly Digest — In Your Inbox

One email a week: top news, releases, and our deepest new guide. No spam. Same content via RSS if you prefer.