GPT-5.6 Model Selection: My Cost-Performance Logic

Stop overthinking the 36+ variants of GPT-5.6. While OpenAI tried to push "auto-routing" to hide the complexity, API users are now staring at a massive matrix of Sol, Terra, and Luna options. If you're trying to optimize your AI workflow without burning your entire budget, you need a mental shortcut.

The community has basically crowdsourced a "民间 (folk) routing guide" because the official docs aren't helping us save money. Here is the breakdown of how to actually pick your model based on the task:

Luna High: This is the sweet spot for daily coding. It's fast, capable enough for most functions, and won't bankrupt you.

Luna XHigh: Use this when Luna High hallucinates or fails a logic check, but you aren't ready to pay the "premium" tax for the top-tier models.

Terra Medium / High: Reserved for repo-level refactoring or building massive new features from scratch.

If you are building an LLM agent for coding, the "effort" setting is the real lever. I've noticed that cranking up the effort on a Luna model often beats using a low-tier Sol model, and it's usually more cost-effective than jumping straight to Terra Ultra.

Since there's no official "cheat sheet," I've been using a system prompt to help my agent decide which model to call via API based on task complexity. Here is a simplified version of the logic I use to categorize tasks:

# Model Routing Logic for Dev Agent
IF task == "small_bug_fix" OR "doc_update" -> USE "Luna High"

IF task == "new_feature_logic" AND complexity == "medium" -> USE "Luna XHigh"

IF task == "system_architecture" OR "cross_file_refactor" -> USE "Terra High"

IF "Luna XHigh" fails twice -> ESCALATE to "Terra Ultra"

The real struggle now isn't whether the model is "smart" enough, but managing the budget. When you have this many tiers, your monthly API spend can spike if your team doesn't have a strict deployment standard.

Basically: start low, push the "effort" slider first, and only hit the Terra series when you're doing heavy lifting. Keep it simple or you'll spend more time managing models than writing code.