GPT-5.6 Luna, Terra or Sol: which should you pick?

Since July 9, OpenAI's GPT-5.6 has come in three tiers, meaning versions of the same model generation sold at different levels of capability and price. Sol is the flagship, Terra is the balanced everyday model, and Luna is the fastest and cheapest. OpenAI says the number marks the generation, while the three names are lasting tiers that can each be updated on their own timetable. So the question "which one?" will come back with every update, and it helps to have a method rather than a favorite.

The short version: Terra is a sensible starting point, Sol earns its higher price on hard, long or specialist work, and Luna is built for volume. The details below show where the gaps between them are small and where they are large.

What separates the three tiers

Developers pay per token, a small chunk of text, often part of a word. Input tokens are what you send; output tokens are what the model writes. Here are OpenAI's published API prices next to three of the benchmark results it reported at launch. A benchmark is a fixed set of test tasks used to compare models.

Two patterns stand out. On coding, the three tiers sit close together, so Terra or even Luna may be good enough for routine programming at a fraction of Sol's cost. On specialized analysis, such as finance research or genomics, Luna falls well behind while Terra stays near Sol. These are OpenAI's own numbers, measured with its own settings, so treat them as a guide to where differences might appear, not as a promise about your tasks.

OpenAI also describes Terra as competitive with GPT-5.5, the previous flagship, at a lower price. TechCrunch, reporting on the launch, summed up the line-up as a workhorse, an intermediate model and a budget option.

A worked task table

The table below maps common jobs to a starting tier. It is a reasonable first guess, built from the pattern above, not a measured result.

In ChatGPT you may not choose the tier yourself in every place. At launch, paid plans use Sol in ordinary chat. Inside ChatGPT Work and Codex, people on the Free and Go plans get Terra, while paid users can pick any tier and an effort level, the setting that controls how long the model thinks before answering. Higher effort usually means more output tokens, so compare tier and effort together rather than one at a time.

When the cheaper tier is the better choice

A cheaper model is not only about saving money. Luna replies faster, which matters in tools people wait on, such as a support chat or an autocomplete. It also lets you afford checks: running Luna twice and comparing, or asking it to classify and then verify, can still cost less than one Sol call. And for very large batches, a small difference in price per item becomes the whole budget.

The reverse also holds. If a cheaper tier gets an answer wrong one time in ten and someone has to find and fix those errors, the time spent can easily outweigh the savings.

How do you test them on your own work?

Build a small test set from real tasks. Pick ten to twenty examples you have already done by hand, so you know what a good answer looks like. Include a few hard ones and a few routine ones, because the tiers differ most on the hard cases.

Run every example through two tiers with exactly the same prompt, files and effort setting, then mark each answer without looking at which tier wrote it. A simple sheet works: one row per task, one column per tier, a pass or fail and a note on what went wrong. Add the time each run took and its token count; in the API, every response reports the tokens it used, which you can multiply by the prices above.

Then read the results as a trade. If Terra passes everything Sol passes, use Terra. If Sol wins only on the hard cases, route those to Sol and the rest to Terra or Luna. Repeat the test whenever OpenAI updates a tier, since the names stay the same while the models behind them can change.