Choosing between GPT-5.6 Soul, Terra, and Luna and just guessing? You’re not alone. I watched a video from an AI creator who’s logged over a thousand hours inside Codex, and he laid out the whole decision tree in a way that finally clicked for me. I’ve been picking models by vibes this entire time. Turns out the “bigger equals better” instinct is just wrong.
Here’s what the expert found, plus the rest of his Codex playbook.
The criteria that actually matter
Before you pick a model, the author says you’re really balancing three things:
- Score: how well the model solves your problem.
- Budget: what you pay per run.
- Speed: how long you’re willing to wait.
He showed a chart plotting score against cost. You want to be high and to the left. High score, low spend. Simple as that. And that’s where the surprise lives.
The comparison
🏆 Soul (largest)
- Pro: highest score across the board. Best for genuinely hard problems.
- Con: most expensive. Overkill for routine work.
Terra (middle)
- Pro: none worth mentioning, honestly.
- Con: scores lower than most Luna variants while costing more. The original poster’s data shows Luna Max beats Terra High on both price and performance. Terra is the trap.
⚡ Luna (smallest)
- Pro: cheapest, and punches way above its size. High or extra-high effort on Luna handles most tasks fine.
- Con: not the one for your gnarliest problems.
Fast mode (any size)
- Pro: roughly 1.5x speedup.
- Con: 2.5x the cost. The math doesn’t work unless you’re truly time-crunched.
The recommendation
The creator’s rule is refreshingly blunt: hardest problems go to Soul, everything else goes to Luna. He sits on Soul High most of the time, drops to medium occasionally, and basically never touches Soul Light. For easier tasks it’s Luna at high or extra-high effort. Skip Terra. Skip Fast unless you’re desperate.
That’s it. Two models, one decision.
Implementation: the setup steps
Once the model choice is sorted, this industry pro walks through the setup that makes Codex genuinely powerful.
- Automate model routing in agents.md You can write rules that pick the model for you. His example: whenever he says “deploy,” Codex spins up a separate project-scoped thread in the same folder using Luna with max thinking. The right model, automatically, every time. No mental overhead.
- Use threads to delegate This one surprised me. In Codex, every thread can see every other thread. So a main thread can spawn a new thread with a specific model and prompt, then manage it. The author demoed it live: he asked his Soul thread to create a Luna Max thread that replies “hello,” and it just worked. The bigger play here is babysitting. Run a bunch of threads in parallel, then spin up one master thread to watch them all. If one stalls, the master notices and nudges it along. That’s orchestration without you sitting there refreshing.
- Audit your agents.md after every model release Your agents.md file collects rules over time. Rules written for older models may be useless or actively harmful now. The expert’s prompt: “Review my agents.md file for any stale rules or things we should revise or remove.” It found plenty of junk in his. Then: “clean it up.” Remember, that file loads into the context window constantly, so bloat costs you on every single call.
- Turn on the browser The person who shared this thinks the Codex browser could become his daily driver by year’s end. It imports your cookies and passwords in one click when you first open it. He’s used it for real work, not just testing: changing DNS records, autoscaling a Supabase instance, reorganizing files into folders. Fast and accurate now, which wasn’t true a while back.
- Install skills Still underrated, per the author. He points to Matt Pocock’s skills collection for coding work. Copy the URL, tell Codex “install this skill,” paste. Done.
- Run loops (called goals in Codex) This is the section I’d steal first. Instead of prompting, waiting 30 minutes, and getting one answer, you give Codex an overarching goal and it keeps working until it hits it. Hours. Sometimes days. His real example: run the benchmark, examine why Astro failed, fix it, repeat until the score hits 90% or above. He kicks it off and walks away. It’s gone from 70% to 90% overnight without him touching anything. He also built a Loop Library and a “Loopy” skill to help you draft loops for your own codebase.
- Control your desktop from your phone Settings, then Connections, then “Control this Mac.” Scan the QR code from Codex mobile. Now your phone drives your desktop’s Codex from anywhere. Since Codex controls your computer, and you control Codex, you effectively control your machine from your pocket. Everything still runs locally.
The safety part you shouldn’t skip
Someone the author knows tweeted that Soul accidentally deleted almost all his Mac’s files. That’s real, and it’s the reason this section exists.
You can’t stop an AI from writing destructive code. Hallucinations happen. What you can do is stop it from running. The fix is lifecycle hooks. The creator set up pre-tool-use hooks that block deletion of the root filesystem, the home directory, and all user directories before Codex executes anything. He shared the prompt as a GitHub gist so you can paste it straight in.
His second piece of advice: don’t run full access. Use “approve for me” instead. A model decides which commands are safe to run silently and flags anything risky for your approval. He runs unrestricted himself but explicitly says most people shouldn’t.
My take
The model comparison alone justifies the watch. Terra being strictly worse than Luna is the kind of thing you’d never guess from the naming. But the loops section is where the real leverage is, and the hooks section is where you avoid a very bad afternoon.
Worth watching the full video for the live demos, especially the thread delegation and the hook config walkthrough.