Anthropic's Claude Code team publishes effort-level tests and recommends low effort for drafting, high for verification

• Levels: low, medium, high, xhigh and max. Opus 4.6 and Sonnet 4.6 have no xhigh, and Claude Code falls back to the nearest lower level (Claude Code model config).
• Defaults in Claude Code: medium on Opus 5.5, Sonnet 5.5 and Haiku 5.5; xhigh on Opus 4.7; high elsewhere.
• How to set it: /effort, --effort, CLAUDE_CODE_EFFORT_LEVEL, or per-model modelSettings.
• Max does not persist: max applies to the current session only unless set through the environment variable. The video says the same (05:19).
• API: set output_config.effort. Per-message effort is in beta (header mid-conversation-output-config-2026-07-01) and keeps the prompt cache (Effort docs).

• What effort governs: all output tokens, including thinking, tool calls and response text. It is a behavioral signal, not a hard token budget. On Opus 5.5, adaptive thinking cannot be disabled, so effort is the main control on cost (Effort docs).
• Benchmark method: 5 attempts per task, 370 attempts in total, GPU tasks excluded, and Fable 5.1 run with production safety interventions off. Production results may therefore differ (Spending your effort).
• Switching cost: changing top-level effort mid-session invalidates the prompt cache. Per-message effort avoids this.
• Re-tuning: the docs say Sonnet 5.5's levels are recalibrated, so settings carried over from Sonnet 5 need a fresh sweep.

• Fitness app, 1.5 vs 67 minutes (low vs max). Medium took 4 minutes and high 11 (Spending your effort). This was a single underspecified build reported by Anthropic.
• 140 vs 214 passes, about 3× tokens. Median was 73k tokens at low and 222k at max (same source). Vendor-run, and not reproduced by the benchmark maintainers. Terminal-Bench 3.0 itself is a community benchmark hosted by Stanford, Harbor and the Laude Institute (tbench.ai).
• 34% → 75%. The Reddit text says this was on "hardest tasks". The video transcript says "hardware tasks" (03:42), and the primary source frames it as a hardware category. Read it as a category result, not a difficulty tier.
• 0/5 → 4/5 on a data-analysis task. Matches Anthropic's gsea-proteomics example.
• A steep time and cost curve is independently supported. TokenSave measured about 42× the time and 11× the cost at max (TokenSave). It did not measure quality.

Effort extends earlier thinking controls. Opus 4.5 paired effort with budget_tokens, and Opus 4.6 and Sonnet 4.6 can still revert to MAX_THINKING_TOKENS (Claude Code model config). Defaults have moved over time:

• Opus 4.7 defaulted to xhigh.
• Opus 5 defaulted to high.
• Opus 5.5 dropped to medium, so a request that omits effort now runs one level lower than it did on Opus 5 (Effort docs).

• Max can backfire. The API docs say it adds significant cost for small gains on most workloads and can cause overthinking on structured-output tasks. Claude Code docs call it prone to diminishing returns (Effort docs).
• Effort does not fix a wrong approach. It reduces missed edge cases, but not a wrong strategy (video 05:39; Spending your effort).
• Defaults differ by surface. The video says Sonnet 5.5 defaults to medium (03:09). That holds in Claude Code, but the API default for Sonnet 5.5 is high.
• The "99%" headline has no supporting data.

• Treat effort as a cost lever. Per-task routing beats a blanket max setting. Hallie's guide also says the default is right for most tasks, and recommends a higher model when Claude fails with full context and higher effort when it skips steps (guide).
• Write a detailed spec first. Anthropic found that detailed specs made results more alike across effort levels (video 06:20), which makes low-effort drafting safer.
• Enterprise admins can cap effort per role (maxEffortLevel) to limit spend.