GPT-5.6 Luna vs Terra vs Sol: Which OpenAI Model Should You Actually Use?

OpenAI sells the same brain in three sizes, and the biggest one costs 25 times more than the smallest. Most people are paying for the wrong one.

Luna, Terra, and Sol launched together on July 9, 2026 as the GPT-5.6 family. The names are pretty; the docs are vague about when to use which. Here's the practical breakdown, with post-price-cut numbers.

Moon, earth, sun: three tiers, one family, a 25x price spread.

The family in one table

Luna: the default nobody admits is enough

After the 80% price cut, Luna is close to free at small scale. And here's the uncomfortable truth for anyone who enjoys paying for the best: for summarization, rewriting, classification, extraction, and everyday chat, Luna's answers are nearly indistinguishable from Sol's. Those tasks don't have enough depth for the extra intelligence to show up. Paying Sol prices for them is renting a supercomputer to run a calculator app.

Terra: where production apps live

Terra is what you pick when Luna starts making mistakes you have to clean up — production code generation, document analysis where a wrong answer costs money, structured multi-step work. At $2/$12 (down 20% in the August repricing), it's the tier most serious apps should build on. The math: Terra costs 10x Luna, so it earns its keep when it saves you more than 10x the cleanup.

Sol: expensive for a reason, unnecessary for most

Sol is the model OpenAI didn't discount, which tells you where the real demand is. It exists for the tasks that genuinely need frontier reasoning: long-horizon agent workflows, hard debugging, research synthesis, the kind of multi-step problems where cheaper models confidently go off a cliff. If your task list doesn't include anything like that, you don't need Sol — full stop.

What about ChatGPT users?

If you only use the ChatGPT app, you never pick a tier directly — the app routes for you. Free users mostly get Luna-class answers; Plus and Pro plans unlock Terra and Sol-class reasoning on harder prompts. Whether that's worth $20/month depends on your usage; our AI subscription guide runs that math.

And the tiering logic isn't OpenAI-specific. Anthropic's Haiku/Sonnet/Opus ladder maps to the same shape, and open-weight models like the best open-source LLMs of 2026 now slot in around Terra quality for a fraction of the cost, if you're willing to do the hosting work.

FAQ

The one-question rule, since I promised it: "If this answer is wrong, what does it cost me?" Nothing much → Luna. Money → Terra. An afternoon of debugging or a client → Sol. Price the failure, not the tokens.