Claude Code Orchestration: 96% Performance at 46% of Cost

Anthropic released first-party benchmark numbers for multi-model orchestration patterns in the Claude lineup. When Fable 5 acts as orchestrator and Sonnet 5 as the worker model, teams get 96% of an all-Fable setup's performance for 46% of the cost (BrowseComp: 86.8% vs 90.8% accuracy, $18.53 vs $40.56 per problem). Flipping the roles — Sonnet as executor consulting a Fable advisor — yields roughly 92% performance at 63% of the cost on SWE-bench Pro.

What makes this notable is that the pattern isn't limited to the API tier: it already works natively for Claude Code subscribers. Three built-in mechanisms enable it — frontmatter in subagent files that pins a cheaper model regardless of the main session, a per-agent 'effort' setting (where low effort on current models reportedly rivals previous-gen xhigh), and a CLAUDE.md policy that tells the main session what to delegate using only role names, so nothing breaks as models rotate.

One caveat worth flagging: since v2.1.198, the built-in Explore subagent inherits the main session's model, meaning background searches can silently bill at Opus rates for daily Opus/Fable users. Defining a user-level Explore agent with model: haiku fixes this. The author packaged the whole setup as 'pilotfish,' a six-role open-source kit on GitHub with one-paste install, a pre-execution plan preview, and a three-step reversible uninstall.

The honest caveat: these are API benchmarks, so subscription savings should track directionally rather than exactly. For zero-setup savings, the built-in /model opusplan command covers most of the same ground.