Releases·Tool Release
Claude Code ships a per-model table of what each effort level costs
A Reddit user found an `effort_cost_index` table inside the Claude Code binary estimating per-model token usage across effort levels, normalized so high = 1. The data shows 'max' means very different costs per model — Opus 5 flattens at the top while Sonnet 5 more than doubles from xhigh to max. The user's own measurements confirmed the direction of the estimates, though magnitudes were loose.
85 engagement1 sourceFri, Sep 18, 2026, 02:08 PM
AI Overview
- An `effort_cost_index` table in the Claude Code binary estimates per-model token use across effort levels, normalized to high=1.
- Sonnet 5 jumps from xhigh to max by +132% (2.41 to 5.59), while Opus 5 rises only +6% (1.60 to 1.70).
- Opus 5's high-to-xhigh step costs +60%, but xhigh-to-max adds just +6%.
- The UI string reads "{from} effort uses more tokens per task than {to}" — consumption per attempt, not per completed or correct task.
- A test on Opus 5 (2 tasks, 4 effort levels) matched the catalog's direction: xhigh and max output tokens measured 397,827 vs 397,761.
- Observed medium-to-high on Opus 5 was x1.17 output tokens vs the catalog's x1.32 prediction; magnitudes ran loose.
- No documentation explains how the catalog numbers were derived or over what workload.
- Fable-5 shows 1.74 at xhigh and 1.91 at max (+10%), while opus-4-8 sits at 1.65 and 1.88 (+14%).
Related
News·Incident
Users report Claude Code CLI and mobile app use different model effort levels, causing unexpected usage spikes
A user discovered that Claude Code's CLI and mobile app can be set to different effort levels (e.g., Opus 5 medium on CLI vs. a higher effort on mobile), leading to rapid consumption of the usage limit when switching devices. This mismatch appears to be a design oversight or bug that can waste users' paid quota.
A user discovered that Claude Code's CLI and mobile app can be set to different effort levels (e.g., Opus 5 medium on CLI vs. a higher effort on mobile), leading to rapid consumption of the usage limit when switching devices. This mismatch appears to be a design oversight or bug that can waste users' paid quota.
48 engagement1 sourceredditSat, Jul 25, 2026, 05:55 AM
Social·Community
User shares Claude Opus 4.8 effort level details from AI team project
A Reddit user building an AI 'team' with Claude shared output from Opus 4.8 Medium detailing the effort ladder (low/medium/high/xhigh/max) and noting that 'xhigh' sits between high and max in cost, while 'max' is a separate 'no constraints' mode. The post provides practical guidance for other users choosing model+effort combinations.
A Reddit user building an AI 'team' with Claude shared output from Opus 4.8 Medium detailing the effort ladder (low/medium/high/xhigh/max) and noting that 'xhigh' sits between high and max in cost, while 'max' is a separate 'no constraints' mode. The post provides practical guidance for other users choosing model+effort combinations.
55 engagement1 sourceredditThu, Jul 23, 2026, 07:01 PM
Social·Community
Users weigh lower effort levels for Claude Opus 5 as cost-saving option
A Reddit user argues that Claude Opus 5's lower effort levels now offer comparable intelligence to higher settings at a fraction of the cost, citing Artificial Analysis scores. The post suggests that as models improve, lower effort levels become more viable for many use cases, potentially changing how practitioners configure cost-performance trade-offs.
A Reddit user argues that Claude Opus 5's lower effort levels now offer comparable intelligence to higher settings at a fraction of the cost, citing Artificial Analysis scores. The post suggests that as models improve, lower effort levels become more viable for many use cases, potentially changing how practitioners configure cost-performance trade-offs.
144 engagement1 sourceredditThu, Aug 6, 2026, 01:02 PM
Developments·Benchmark
Claude Opus 5 vs Opus 5.5 Effort Levels Benchmark
An independent benchmark by George Liu runs 120 headless Claude Code prompts against Claude Opus 5.5 and Opus 5, testing all five effort levels (low, medium, high, xhigh, max). The results give practitioners a like-for-like view of how the newer model and effort settings affect real coding-task performance.
An independent benchmark by George Liu runs 120 headless Claude Code prompts against Claude Opus 5.5 and Opus 5, testing all five effort levels (low, medium, high, xhigh, max). The results give practitioners a like-for-like view of how the newer model and effort settings affect real coding-task performance.
6 engagement1 sourceredditSat, Sep 26, 2026, 12:34 PM
Products·launched
effort-cycle: Claude Code plugin to switch effort levels with Alt+E and a footer meter
A Claude Code plugin (hooks-based mod) that lets developers cycle reasoning effort levels with Alt+E / Alt+Shift+E instead of typing /effort. The footer displays the current level as a meter next to the model name. Built by the author using Claude Code itself.
A Claude Code plugin (hooks-based mod) that lets developers cycle reasoning effort levels with Alt+E / Alt+Shift+E instead of typing /effort. The footer displays the current level as a meter next to the model name. Built by the author using Claude Code itself.
3 engagement1 sourceredditMon, Sep 28, 2026, 09:42 PM