Max out your Claude subscription: Opus orchestrates, cheap models execute tickets
EXM7777 · x · 2026-09-26
A practical tip for stretching your Claude subscription: run subagents in parallel with smart tiering:
- One main chat on Opus 5.5 xhigh acts as the orchestrator
- Plan the session as tickets using Matt Pocock's skills
- Each subagent gets one ticket plus the smallest context it needs
- Complex tickets go to Opus 5.5 medium; everything else to a really cheap model
Crucially, the executing model doesn't have to be Claude — Claude Code can call another harness as a headless worker. The author uses pi․dev with Deepseek V4.1 flash, or Codex with GPT-6 Luna Fast.
More from coding & agent
- AI dev discourse has shifted from frameworks to model-version semantics, says engineer — vboykis · 2026-09-26
- Devs mock coding models that ignore better-auth and hand-roll custom auth — haydendevs · 2026-09-26
- Open-source Minara AI CLI runs autonomous agent workflows via MCP from your desktop — dr_cintas · 2026-09-26
- Codex merges Memory V2 prompts: 10KB summary cap, citations and selective retrieval — morqon · 2026-09-26
- AgentOdyssey, a Text-Game Engine for Test-Time Continual Learning Agents, Accepted at NeurIPS — DanielKhashabi · 2026-09-26
- Mario design rules for vertical agents: consistent inputs and progressively unlocked capabilities — dragon_khoi · 2026-09-26