[Claude Code] Dividing labor among models used about 2.7 times the tokens of a single Opus

Agents & Dev✍ TextPublished: Source: Zenn (LLM)

#Anthropic #Claude Code

Key point

When dividing article production into four roles among Haiku, Sonnet, and Opus, it used about 2.7 times the tokens of a single Opus. The single-model version was also superior in terms of fewer errors and information volume.

According to a report on Zenn, when dividing article production into four roles among Haiku, Sonnet, and Opus using Claude Code, it consumed approximately 2.7 times the tokens used by a single Opus. This comparison is based on a single actual measurement where all models were from the 5.5 series.

The report states that the single-model version was superior in both the number of fewer errors and the volume of information. It also noted that even when a verifier is placed, they could not notice "information that was not collected" that they had not been asked to confirm.

Since switching to cheaper models does not necessarily mean token usage will decrease, it is recommended to decide after measuring compared to when using a single model. Additionally, it was reported that four sub-agents could be operated under the instructions of Opus.

This item is a summary. The full, checkable version is at the original source:

Read the original at Zenn (LLM) ↗