Engineering · 7 min read · April 19, 2026
Claude Code model tiers and effort levels, explained plainly
Choosing the wrong model or effort level in Claude Code wastes tokens silently. Here is what each setting actually controls.
Claude Code offers three model tiers and four effort levels that independently control reasoning depth and cost, but most developers never adjust either.
- — Haiku suits fast lookups and mechanical tasks; Sonnet handles most daily coding work.
- — Opus unlocks 1M token context on Max, Team, and Enterprise plans automatically.
- — Plan mode enforces read-only access at the tool level, not just as a soft guideline.
- — The opusplan alias switches from Opus to Sonnet at execution time but loses the 1M context window.
- — Effort levels (low, medium, high, max) control internal reasoning token budget before output.
- — Anthropic changed Opus 4.6 default effort from high to medium in v2.1.68, causing community friction.
- — Subagents inherit the parent session model and effort unless explicitly overridden in frontmatter.
- — Higher effort can cause overthinking on simple tasks, producing verbose output where directness was needed.
Frequently asked
- Model tiers determine the underlying capability of the model being used. Haiku is the fastest and least capable, Sonnet is the general-purpose default, and Opus is the most capable for complex reasoning. Effort levels are a separate control that governs how many internal reasoning tokens the model spends before producing a response. Low effort is fast and minimal; max effort removes the reasoning token ceiling entirely. The two settings are independent, so you can run Opus at low effort or Sonnet at high effort depending on the task.