Claude Code ships a per-model table of what each effort level costs

A hidden table within the Claude Code binary, named effort_cost_index, reveals per-model token-usage estimates for various effort levels. Normalized with 'high' set to 1, the data shows sonnet-5 has a 'max' cost 132% higher than 'high', opus-4-8 is +14%, opus-5 is +6%, and fable-5 is +10%. The source notes that the catalog numbers' derivation and workload are undocumented, suggesting they be treated as vendor estimates.

This is the first time a per-model token-usage estimate for different effort levels has been extracted directly from the Claude Code binary, unlike public documentation.