An individual XDA experiment applied Spotify’s shunt approach to Claude Code and other task models but did not achieve the expected savings. Opus 5.5 already handled large files through search and partial reads, so the read hook rarely activated. For test generation, a standalone run used 31,900 message tokens and took about one minute, while delegation to local Qwen used 39,300 tokens and delegation to Haiku used 42,800, with delegated runs taking 7–10 minutes. Reading and correcting errors in drafts increased the cost. The number of generated tests also differed, so this was not a controlled comparison of general performance. Delegation and verification costs need to be measured together on one’s own workloads.
Back to Tech
Claude Code delegation experiment increases tokens by 23–34% instead of saving them
A tool that reduces large-file reads can still cost more, depending on existing reading behavior and the effort to repair delegated output.