Delegating to cheaper models: a 3-codebase test shows up to 75% savings with quality trade-offs
BeginningSir7334 · reddit · 2026-11-04
The author built a model-router skill letting Claude Opus delegate to Haiku/Sonnet, then compared four setups (forced Opus, spontaneous Opus, skill-routed, forced Haiku) across three codebases (10k, 22k, 52k lines).
Results
- Small codebase: all setups passed 10 everyday tasks; forced Opus was most accurate per run, skill-routing cost 30% more.
- Medium codebase: forced Haiku found 11/12 known doc issues, matching Opus at 41% less cost; skill-routing cost 11% more than forced Opus.
- Large codebase: biggest savings — spontaneous Opus 32% cheaper, skill-routed 38%, forced Haiku 75%. But on an 80-page doc check, forced Opus was clearly most accurate (45/51 issues), forced Haiku found 34 at 84% less cost.
Mechanics: small projects favor Opus writing inspection scripts (delegation adds briefing/verification overhead); medium projects favor Haiku's cheap cached-context reading; large projects benefit from splitting long expensive sessions into smaller parallel contexts, with accuracy trade-offs.
Key observations: re-pricing delegated tokens at Opus rates misleadingly suggests 34–85% savings, but head-to-head results ranged from 30% more expensive to 75% cheaper. Opus never delegated spontaneously in any run. No universal best approach — the choice depends on codebase size and task type.
More from coding & agent
- LangChain benchmarks 7 browser toolsets: Stagehand wins with 90% pass rate at $1.98 vs $83.62 — LangChain · 2026-11-04
- Xogot ships built-in Pi.dev-powered coding agent for Godot on iPad — mitsuhiko · 2026-11-04
- DeepSeek quietly launches "DeepSeek Harness" YouTube channel, hinting at plugin-based agent framework — gnukeith · 2026-11-04
- Sharing agent sessions across a team breaks knowledge silos — and AI can summarize your teammates' workflows — steipete · 2026-11-04
- Apgar: open-source release health checker catches agent regressions without an LLM judge — krstCB · 2026-11-04
- Claude-based agent transcribes entire site with images into context for Claude 3 Opus — repligate · 2026-11-04