A developer-shared configuration for Anthropic's Claude Code reportedly routes one coding session across three models at once, with the stated aim of producing better output while consuming fewer tokens.
Claude Code · Developer configuration
One coding session, three models: a developer's split-workload setup
A community-shared configuration reportedly launches Claude Code with one command and divides work across Sonnet 5.5, Haiku 5.5 and Opus. The stated goal is better output with fewer tokens. The flags and model roles have not been confirmed by Anthropic.
3
models running in a single session
1
launch command: claude --advisor opus --subagents haiku
0
official confirmations of the flags or their behavior reviewed
How the work is divided
Each model takes one stage of the coding task.
Advisor
Opus
Provides guidance. How it interacts with the other two models is not specified.
→
Main session
Sonnet 5.5
Writes the code and handles the core reasoning.
←
Subagents
Haiku 5.5
Explores files and documentation in parallel, a lighter and faster read-through.
Why it could work
Heavy reading of a codebase goes to a cheaper model, leaving the main model's context for writing.
Parallel exploration may speed up the search-and-read stage.
Separating reading, writing and review mirrors how many developers already divide work.
What remains open
The claim of better output is contestable and depends on how well the models coordinate.
No benchmarks, pricing comparisons or independent tests of token savings were reviewed.
The flags and model names have not been independently verified.
Outlook
Multi-model orchestration may become standard, but only if vendors formalize it.
• Whether official support is added to Claude Code
• How reliably context is handed off between models
• Whether measured token savings hold across real projects
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...