Claude Code Reveals Effort Level Cost Metrics by Model

Data named effort_cost_index was discovered inside the Claude Code binary, revealing the relative token usage for Effort levels (low, medium, high, xhigh, max) by model. This metric normalizes the high level to 1.0, allowing cost comparisons across other levels.

Differences in Cost Structure by Model

When performing the same task, the rate of cost increase with rising Effort levels varies significantly across models.

• Opus 5: Moving from high to xhigh increases costs by 60%, but moving from xhigh to max increases costs by only 6%. This shows a trend of diminishing cost increases at the top levels.
• Sonnet 5: Moving from high to xhigh increases costs by 141%, making the max level effectively the most expensive tier.
• Fable 5: Moving from xhigh to max increases costs by 10%.

Limitations of Measurement and Verification

This metric represents 'token consumption per task attempt' and does not include task success rates or retry costs. The author measured results by running 8 agents with Opus 5; while the directional prediction (cost similarity between xhigh and max) was accurate, there were deviations in absolute figures. In particular, a high cache read ratio limited total token count measurement, so this should be considered a vendor estimate.