Frugal routes Claude Code sub-tasks to cheaper models and escalates only on real failures

StolenDrinks · reddit · 2026-07-23

Frugal routes Claude Code sub-tasks to the cheapest model that can finish them

The author shares Frugal, a Claude Code plugin that sends each sub-task to the cheapest capable model and escalates only when a real check fails.

Routing ladder

Deterministic shell tools like grep, jq, and git if they can answer the question.
A Haiku worker for locating and extracting information.
A Sonnet worker for mechanical edits from a spec.
The main model for design, debugging, and review.
A top-tier model only as the final escalation ceiling.

Design choices

Escalation is verified, not self-reported. Tests, compiler checks, schema validation, or patch application failures trigger a step up.
Enforcement is via hooks, not prompts. Two PreToolUse hooks count search calls and can hard-block the reasoning tier at spawn.
Context stays isolated. Instead of swapping models mid-session, Frugal spins up subagents with only pointers such as file paths, line ranges, SHA values, and short specs.

What it reports

Local logs for each worker run.
Per-tier spend, escalation rate, and savings versus using the session model for everything.
No telemetry.

The repo is explicitly Claude Code-only and the author asks for critique on the escalation heuristic and routing table.

More from coding & agent