Effort Levels — The Reasoning Slider
Effort levels are a session-wide reasoning depth control — separate from ultrathink's per-prompt override. Opus 4.7 introduced xhigh as the new default, sitting between high and max. Understanding the five levels prevents both under-spending on hard problems and over-spending on easy ones.
Quick Reference
- →Five levels: low, medium, high, xhigh, max
- →xhigh: new tier in Opus 4.7 — Opus 4.7 defaults to xhigh, not high
- →Opus 4.6 / Sonnet 4.6 on Max/Enterprise/Team: defaults to high
- →Opus 4.6 / Sonnet 4.6 on Pro: defaults to medium
- →Other models silently fall back to high if xhigh is requested but not supported
- →/effort command: interactive arrow-key slider (April 2026) — no argument needed
- →CLAUDE_CODE_EFFORT_LEVEL env var: persistent default across sessions
- →Effort is model-calibrated: high on Haiku ≠ high on Opus 4.7
The Five Effort Levels
Effort levels control reasoning depth, agentic persistence, and how aggressively Claude pursues tool calls before responding. They are not fixed token budgets — they are calibrated to each model's capacity and scale dynamically based on the complexity of the current turn.
| Level | Symbol | Default for | Use case |
|---|---|---|---|
| low | ○ | — | High-volume tasks, quick lookups, simple classification — speed and cost over depth |
| medium | ◐ | Opus 4.6/Sonnet 4.6 on Pro | Balanced performance — solid results without full token expenditure |
| high | ● | Opus 4.6/Sonnet 4.6 on Max/Team/Enterprise | Deep reasoning for complex tasks — the baseline for serious implementation work |
| xhigh | ◉ | Opus 4.7 (all plans) | Extended exploration, repeated tool calling, complex agentic coding — sweet spot between performance and cost |
| max | ⬭ | — | Maximum reasoning, no token constraint — only for the hardest architectural decisions |
xhigh was introduced with Opus 4.7 in April 2026. It sits between high and max, and is specifically designed for the extended agentic work that Opus 4.7 excels at — complex multi-step coding, repeated tool calling, exploration before implementation. Anthropic describes it as 'the compelling sweet spot between performance and token expenditure.'