Strong model plans, cheap models build: open-weight coders now good enough to make dev nearly free
Feisty-Sky-6018 · reddit · 2026-09-12
The author shares a workflow where a strong model plans and reviews while cheap/open-weight models implement — and argues that with a good enough spec, the builder model matters less than ever.
- Following Anthropic's long-running-agent approach, Fable 5.1 generates a spec. with small, verifiable tasks including context, constraints and explicit verification
- Builders tested include Laguna S 2.1, Qwen, Kimi K2.7 Code, GLM-5.3, Haiku and GPT-5.4 nano; once the spec is good, output quality converges (mostly 7-8/10), with speed being the main differentiator
- Loop: strong model plans → cheap model does 5-10 tasks → early review → fix/split/reorder → repeat; plus a Ralph Wiggum-style shell loop and optional Solvr A2A rooms for complex tasks
- Takeaways: spend expensive tokens on thinking, cheap ones on execution, keep tasks small, review early, keep context in files; open-weight models are already good enough for most implementation work
- The experiment project, spec and quota-tracking script (which tells the author which provider to use) are open-sourced on GitHub
More from coding & agent
- Codex app 26.908 ships Quick Chat from Pets controls and Win+Alt+P shortcut on Windows — TheMoonMidas · 2026-09-12
- AI ports a browser three.js demo to Unreal Engine 5.8 without Blender — TheMoonMidas · 2026-09-12
- Anthropic's Jacalazu shares her agent-minutes formula for maximizing AI agent output — jacalulu · 2026-09-12
- Dev splits workflow: local scheduled tasks on Codex, cloud tasks ported to Grok Bot — petergyang · 2026-09-12
- OpenAI ships Docs skill for Codex plus a Developers plugin for building and managing agents — gabrielchua · 2026-09-12
- OpenAI Agents API ships with sandboxes, compaction, and multi-agent support — gabrielchua · 2026-09-12