Claude Code Advisor Call Cost Flow

Sonnet 5 executor + Opus 5 advisor: where the tokens and dollars go (measured, Claude Code v2.1.226)

Claude Code Advisor Call Cost Flow Sonnet 5 executor + Opus 5 advisor: where the tokens and dollars go (measured, Claude Code v2.1.226) claude -p --advisor claude-opus-5 write code, run tests test results (iterate) advisor call: full transcript 53,210 uncached input tokens (Exp 1) - cache_read 0 advisory response 2,828 output tokens - $0.337 for the call apply advice, continue final result (JSON) executor row at Sonnet rates separate advisor row total cost = sum of both rows Execute at Sonnet rates Advisor call (not guaranteed) Cost accounting Engineer · claude -p CLI · Sequence participant Engineer claude -p CLI Sonnet 5 · executor model · Sequence participant Sonnet 5 executor model Workspace · code + tests · Sequence participant Workspace code + tests Opus 5 · advisor_20260301 · Sequence participant Opus 5 advisor_20260301 modelUsage · JSON cost report · Sequence participant modelUsage JSON cost report Legend request return async trace default message

Two Opposing Forces

  • • Savings: most output is generated at Sonnet rates
  • • Overhead: every call re-reads the whole transcript uncached
  • • Break-even is a race of absolute amounts, not ratios

Uncached Input Dominates

  • • Advisor-side cache_read was 0 on every call
  • • Per-call input grew 53,210 to 77,611 tokens as the chat grew
  • • The advisory response itself stays small (2,828 output tokens)

Measured Outcomes

  • • Short task, 1 call: +25% vs fixed Opus ($1.238 vs $0.993)
  • • 435-line refactor, 2 calls: -15% ($3.094 vs $3.646)
  • • No calls at all: 43% cheaper than fixed Opus

Sequence diagram • Built with Archify • Create yours ↗ • Hover to trace • R route • Click to focus • +/− zoom • M radar • [/] views • P play story • T theme • E export