Stop reaching for the biggest model: a cost-efficient Cursor/Codex setup with GPT-6.1 Sol
chiliraupe · reddit · 2026-10-02
The author shares their Cursor/Codex configuration for consulting and automation work, arguing that top-tier models like Sonnet/Opus/Astra are overkill for most daily tasks.
Current setup
- Main agent: GPT-6.1 Sol (Medium)
- Subagent: GPT-6 Luna (Medium)
Findings
- GPT-6.1 Sol reaches 86% of Opus's score at roughly 36x lower cost
- GPT-6 Luna is nearly free and works well as a subagent for research, file analysis, and other subtasks
- Daily uses include Power Automate flows, Plane project management, ticket/PRD writing, process analysis, documentation, AI consulting, and MCP/tooling config — no case yet requiring a much larger model
The takeaway: benchmark points and productivity aren't the same thing; performance-to-cost ratio should drive model choice, making this combo a sweet spot for consulting and automation workflows.
More from coding & agent
- AI writes more code than you can read: a senior engineer's 6-step review path — MaryamMiradi · 2026-10-02
- Dev building GTA 4-style SCP game entirely with Claude Opus 5.5 shares first demo — imjustnewatai · 2026-10-02
- Claude Code Mods: spin off forked agents to build your own memory harness — trq212 · 2026-10-02
- The Hard Part of AI Memory Is the Write, Not the Read — Tiwaryswarnim · 2026-10-02
- Codeck presentation tool fully integrated inside Codex as a plugin, runs live prompts on slides — Dimillian · 2026-10-02
- Voice-controlled AI agent unlocks developer's apartment on command — Baconbrix · 2026-10-02