GPT-5.6 Sol, Terra and Luna: What Creators Actually Need to Know

Three Names, One Family

Somewhere in your studio there’s a note that says “Thinking for briefs, Instant for captions”. You can throw it away. OpenAI has reorganised the whole thing, and the new shape is worth ten minutes of your attention because you’ll be living in it for a while.

GPT-5.6 arrived in preview on 26 June and went on general release on 9 July. It comes in three tiers, named after the sun, the earth and the moon, which is at least easier to remember than a string of decimals. Sol is the flagship: the smartest, the slowest, the one you want for anything hard. Terra is the balanced middle, priced at roughly US$2 per million input tokens and US$12 per million output on the API. Luna is the fast, cheap one at US$0.20/US$1.20 — the model for captions, alt text and the two hundred variations of a subject line you need by lunch. All three carry a one-million-token context window, so you can hand any of them a whole script, a brand book, or a season of transcripts and it will keep it all in view.

This replaces what I wrote earlier this year about GPT-5.5. The image side of that story survives largely intact. The model side does not.

Which One You’re Actually Using

Here’s the bit OpenAI doesn’t put on the pricing page in large type. Which tier you get by default depends on what you pay.

• Free and Go. Luna has been the default since 6 August. It’s quick and perfectly good for light work, but if you’ve noticed the free ChatGPT feeling a touch shallower on a complex brief lately, that’s why.
• Plus and Pro. Sol is the default. This is the model most of the impressive screenshots come from.
• Terra. The middle sibling is mostly a choice you make on purpose — through the API, or by switching manually — rather than one you land on by default.

If you’re a freelancer on the free plan wondering why your colleague’s results look better, the answer is usually not their prompting. It’s that they’re talking to a different model. Worth knowing before you spend a weekend rewriting your prompts.

The Death of Instant and Thinking

In the old ChatGPT you picked a mode. Instant answered quickly; Thinking went away for a minute and came back with something more considered. With GPT-5.6 both are folded into one model with a reasoning-effort slider. Same brain, adjustable patience.

Practically, for creative work:

• Low effort is the old Instant. Captions, hashtags, quick rewrites, “give me twenty names for this”. It’s cheap and it doesn’t overthink, which for brainstorming is actually a virtue — the fast answers are looser.
• Medium is where I leave it most of the day. Treatments, shot lists, first-draft copy.
• High effort is the old Thinking. Use it when getting it right first time is worth waiting a minute: a structured pitch document, a script with a real logic problem in it, a brief you’ll be judged on. It costs more tokens and takes longer, so don’t leave it there for everything or you’ll burn through your allowance answering “is this comma right”.

The habit to build is simple. Ask yourself whether you want the answer fast or right, and move the slider accordingly. That’s the whole skill.

ChatGPT Images 2.0: It Thinks Before It Draws

ChatGPT Images 2.0 has been live on every plan since late April, and it’s the piece of this that changes day-to-day creative work the most. The API name is gpt-image-2, if you build things.

The important difference from every image model you’ve used before is that it reasons before generating. Ask for “a poster with three people, two dogs and the headline in the top third” and it plans the composition and checks the counts and constraints before it commits pixels. It’s the difference between a junior who starts drawing the moment you finish speaking and one who reads the brief back to you first. Add up-to-2K output, up to eight images per prompt, and text rendering that finally handles a real headline — in several languages — without turning it into runes, and you have something a working designer can rely on rather than gamble on.

One catch. There are two modes. Instant is free for everyone. Thinking — which adds web search, layout reasoning, multi-image batching and a verification pass — is Plus, Pro, Business and Enterprise only. If you’re on the free plan, you get the better model but not its better habits.

If you’re a designer

Use the eight-per-prompt batching for what it’s good at: exploring a direction, not finishing one. Ask for eight compositions of the same brief, pick two, then iterate on those individually. Put the actual headline in the prompt now that text works — lorem ipsum is no longer necessary and it changes how you read the layout. And keep your own type on top for anything going to print; “good text rendering” is not the same as kerning.

If you’re a filmmaker

This is a previs and pitch tool. Storyboard frames, lookbook pages, a consistent character across eight angles for a deck. The reasoning step makes it much better at holding a described shot — lens, height, light direction — than the old model was. It is not a video model; for motion you’re still in Kling, Veo, Seedance or Omni Flash territory. But the frames you take into those tools will be better if you plan them here first.

If you’re a marketer

The counts-and-constraints checking is your friend. “Five product shots, white background, the logo bottom-left, no text other than the price” now mostly arrives as ordered. Thinking mode’s web search is genuinely useful for “in the style of our current campaign” work when the campaign is public — and a reason to be on a paid plan if your team makes a lot of social variants.

Against Midjourney V8.2 and Nano Banana

I keep a longer, fairer three-way comparison up to date, so here’s the short version.

Midjourney V8.2 is still the one for beauty. If the job is a mood, a texture, a painterly frame that makes a client go quiet, Midjourney remains hard to beat, and since V8.1 it’s four to five times faster at 2K, which removed its biggest practical complaint. It is worse than Images 2.0 at doing precisely what you said.

Google’s Nano Banana 2 outputs at 4K, which matters if the work is going large, and Nano Banana Pro is the strongest of the lot at editing an existing image without wrecking the parts you liked. The Lite version turns an image round in about four seconds, which is what you want for a live client session.

Images 2.0 wins on instruction-following, text, and consistent batches. It’s the one I’d hand a brief with numbers in it. Most working creatives I know now run two of the three and don’t apologise for it. That’s not indecision; it’s a kit.

Building Tools, and Who Owns the Output

If you build creative tools rather than just use them — a caption generator for a client, a storyboard app, an internal batch pipeline — the API pricing has moved in your favour. Sol dropped from US$5/US$30 to a promotional US$4/US$20 per million tokens, running until at least 21 November 2026. Terra at roughly US$2/US$12 is the sensible default for most creative-tool backends, and Luna at US$0.20/US$1.20 makes high-volume, low-stakes generation almost free. Microsoft 365 Copilot now offers GPT-5.6 with effort levels too, so if your clients live in Word and PowerPoint they’re getting the same models under a different roof.

On ownership: OpenAI’s terms assign to you whatever rights it holds in what the model produces. So as far as OpenAI is concerned, the poster is yours to sell. Whether an AI-generated image attracts copyright protection at all is a different question, answered differently in different countries and still being argued in courts — the same week as this post, the US Department of Justice was in court defending training on copyrighted text as fair use while music publishers were suing Anthropic. My working rule for client jobs hasn’t changed: treat AI output as an ingredient in a human-authored piece, not as a finished asset you can register, and say so in the contract.

The Bottom Line

Know which tier you’re on. Learn the slider — fast or right, that’s the question. Brief Images 2.0 like a person, in a list, and let it batch. Keep Midjourney for beauty and Nano Banana for edits. If you build, look at Terra.

None of that is dramatic, and that’s the point. The tools have settled into a shape you can actually work in. The interesting question now isn’t which model is best. It’s what you’ll make with the afternoon you just got back.

Frequently Asked Questions

What is the difference between GPT-5.6 Sol, Terra and Luna?

They’re three tiers of the same model family, all with a one-million-token context window. Sol is the flagship and the default on ChatGPT Plus and Pro. Terra is the balanced middle tier. Luna is the fast, cheap tier and has been the default on the free and Go plans since 6 August 2026. On the API, Sol is US$4 per million input tokens and US$20 per million output on a promotional price running to at least 21 November 2026, Terra is around US$2/US$12, and Luna is US$0.20/US$1.20.

What happened to ChatGPT Instant and Thinking modes?

They’re gone. With GPT-5.6, ChatGPT folds both into one model with a reasoning-effort slider. Low effort behaves like the old Instant mode and is right for captions, quick rewrites and brainstorming. High effort behaves like the old Thinking mode and is worth it for briefs, structured scripts, and anything where you’d rather wait a minute than fix the answer.

What does ChatGPT Images 2.0 do that the old image model didn’t?

It reasons before it draws — planning the composition and checking counts and constraints in your prompt — and it outputs at up to 2K resolution, up to eight images per prompt, with much stronger text rendering including non-English scripts. It’s been live on all ChatGPT plans since late April 2026. Its Thinking mode, which adds web search, layout reasoning, multi-image batching and verification, is only on Plus, Pro, Business and Enterprise; Instant mode is free.

Is ChatGPT Images 2.0 better than Midjourney or Nano Banana?

It depends on the job. Images 2.0 is the strongest of the three at following complicated instructions, rendering text, and producing consistent batches. Midjourney V8.2 still leads for aesthetic and painterly work and is far faster since V8.1. Google’s Nano Banana 2 outputs at 4K, and Nano Banana Pro is the strongest for precise edits of an existing image. Most working creators use more than one.

Who owns the images and text I make with GPT-5.6?

OpenAI’s terms assign to you whatever rights it holds in the output, so as far as OpenAI is concerned the work is yours to use commercially. Whether AI-generated output attracts copyright protection at all is a separate legal question that varies by country and is still being settled, so for client work treat AI output as an ingredient in a human-authored piece rather than a finished asset you can register.