GPT-5.6 Is Here: Sol, Terra, and Luna Explained for Everyday Users

OpenAI shipped GPT-5.6 on Friday, June 26, 2026, and in the same breath retired GPT-4.5, the model a lot of people still had saved workflows built around. The new family arrives under initial restrictions rather than as a switch everyone can flip, according to its Wikipedia entry, and it comes in four variants with names that sound like a tarot deck: Sol, Sol Ultra, Terra, and Luna.

If you use ChatGPT to get through a workday rather than to follow AI news, you have two reasonable questions. What are these names, and what should I actually click? The second question turns out to have a cleaner answer than the first.

The picker changed before the model did

The bigger change for daily use actually landed two weeks earlier. On June 10, per OpenAI’s release notes, ChatGPT’s model picker was simplified to five options: Instant, Medium, High, Extra High, and Pro.

Read those again. Not model names. Effort levels.

This is a real philosophical shift, and a merciful one. The old picker asked a regular user to hold a mental map of version numbers and suffixes that changed every few months, a chore roughly nobody signed up for. The new picker asks one question you can actually answer: how hard should the machine think about this? You choose the effort; ChatGPT sorts out which model does the work underneath. The model names become plumbing.

Two retirements completed the cleanup. GPT-5.2 went on June 12 and GPT-4.5 on June 26, both per the same release notes. The lineup behind the picker got newer while the picker itself got simpler, which is the right direction for both.

So what are Sol, Terra, and Luna?

Here is the honest answer, which is shorter than the confident answers you will find elsewhere: GPT-5.6 has four variants, named Sol, Sol Ultra, Terra, and Luna, per Wikipedia’s entry on the release. That naming, sun and earth and moon with an Ultra on top, replaces the mini-and-nano suffix convention OpenAI used for years.

What OpenAI has not done, as of this writing, is publish the kind of detailed public breakdown that would let anyone say with a straight face what Sol is best at or how Terra differs from Luna in practice. The release is two days old and restricted. Anyone currently ranking the variants for you is extrapolating from the names, which is astrology, fittingly enough.

I wrote a first draft of this section that guessed at the tier structure, and deleted it, because guessing dressed up as explaining is most of what is wrong with AI coverage. When OpenAI publishes specifics, this site will cover the specifics. Until then: four names, one family, limited availability, and a picker that mostly hides the question anyway.

Which effort level for which work

Since the picker is what you actually touch, here is a working rule of thumb for matching its five levels to office tasks. This part is advice, mine, not OpenAI documentation.

Instant is for the throwaway tier: rewording a sentence, a quick factual lookup, “make this email 30% shorter.” Speed matters, stakes are low, and waiting on deep reasoning to fix a typo is silly.

Medium is the everyday default, and unless a task is clearly above or below it, staying here is correct. Summaries, standard drafts, meeting notes, the bulk of a normal day.

High is for work with structure: a document where the logic has to hold together, analysis of a messy spreadsheet, a first pass at a plan someone will poke holes in.

Extra High is for the tasks you would give your sharpest colleague and a full afternoon: hard tradeoffs, long multi-part documents, anything where a subtle error costs real money or credibility. Slower, and worth it precisely when it is worth it.

Pro sits at the top for the heaviest reasoning work, which for most office users is an occasional need, not a daily one.

In table form, since that is how most people will actually remember it:

Picker level	Reach for it when
Instant	Quick rewrites, lookups, low-stakes edits
Medium	Normal daily work; the default for a reason
High	Structured documents, analysis, plans
Extra High	High-stakes work where subtle errors cost money
Pro	The rare, genuinely hard reasoning problem

The mapping between these levels and the models underneath is OpenAI’s business, and it will keep changing as models come and go. That is the entire appeal of the design: the last three weeks saw two retirements and a major release, and a user who only knows the five levels felt none of it.

A cheap habit that beats memorizing any of this: if a Medium answer feels shallow, re-run the same prompt one level up and compare. Thirty seconds, and you learn where your own tasks actually sit on the ladder. Better prompts move results more than the picker does, though, and that skill is covered in our guide to writing AI prompts.

Retired models and your saved stuff

The part of a model retirement that worries regular users is the archive. Two years of chats, saved prompts taped inside a team wiki, that one thread where ChatGPT finally understood your invoicing format.

The calming news: retirement removes a model, not your data. Chat history stays. Continue an old GPT-4.5 conversation and it simply picks up on a current model. Nothing you wrote disappears on June 26.

The catch is subtler. A new model reading an old thread is a new employee handed a binder: they have the notes, not the habits. Responses in continued chats may differ in tone, format, or judgment from what the retired model produced. For casual chats, irrelevant. For workflows, it means one thing: re-test before you trust. Take every saved prompt your team actually depends on, run it fresh, and check the output against what you expected. Most will be fine or better. The ones that break, you want to find on a Tuesday morning, not inside a client deadline.

Why do vendors keep doing this? Running a model costs money every hour it stays available, and every old version kept alive is an old version to maintain, secure, and support. Retiring superseded models is how a lab keeps its lineup affordable and its engineers pointed forward. That does not make the churn painless for the person whose favorite model just vanished mid-quarter, but it does mean the churn is structural. Plan for it the way you plan for software updates generally: assume the ground moves about twice a year, and keep your important prompts written down somewhere outside the chat history so re-testing them takes minutes instead of an afternoon of archaeology.

If this churn is wearing on you, you are not imagining it; the site’s Claude vs ChatGPT comparison got rewritten this month for the same reason, and the alternatives roundup exists for people ready to shop around. Every vendor does this now. Model retirement is a recurring weather pattern, not an OpenAI quirk.

What to do while the restrictions last

Nothing dramatic. GPT-5.6 launched under restrictions, no full-rollout date has been announced as of June 28, and refreshing the app will not speed that up. The useful moves this week are smaller: get comfortable with the five-level picker, since that interface is where you live now. Re-test your saved prompts on the current models, since the retirements already happened whether you noticed or not. And when GPT-5.6 does reach your account, resist the urge to rank Sol against Luna based on vibes and a name. Run your own work through it. A model’s real spec sheet is what it does to your Thursday.