AI
[Workflow] Claude Code Cost Optimization: Understanding Billing Mechanics and Using Pilotfish for Smart Model Delegation
A new community workflow and open-source tool named "pilotfish" has been released to optimize Claude Code billing.
Key takeaways
- The built-in Explore subagent bills at the main session's model tier since v2.1.198, accounting for roughly 36% of session API calls.
- Implementing the pilotfish multi-agent delegation system reduced costs from $14.50 to $3.70 in a 12-worker audit.
- The pilotfish installation uses a 'curl | sh' style script which requires users to review the code first for safety.
A new community workflow and open-source tool named "pilotfish" has been released to optimize Claude Code billing. The strategy addresses high API costs caused by hidden subagent model inheritance - such as the built-in Explore subagent billing at the main session's model tier since version 2.1.198. By delegating tasks to cheaper models like Haiku for scouting and Sonnet for mechanical work, a measured audit showed a cost reduction from $14.50 to $3.70.
It provides a concrete method to resolve unexpected and high billing costs in Claude Code.
By the numbers
- 36%
- of a session's API calls going through Explore
- $14.50
- cost of all-Fable session before optimization
- $3.70
- optimized cost using Fable orchestrating Haiku workers
How it unfolded
- Outage served as natural experiment for fallbackModel behavior
Turn stories like this into views
Ravenclip finds the AI news, makes the video, and posts it before attention moves on.
Common questions
- What happened with Claude Code?
- The built-in Explore subagent bills at the main session's model tier since v2.1.198, accounting for roughly 36% of session API calls.
- Where can I read the original report?
- Read the full report at Reddit Search.