OpenAI's models: from GPT-5.6 to GPT-6, Sol, Terra, Luna and Codex

GPT-5.6 isn't one model, it's a generation. The number 5.6 only shows when it was born, not how strong it is - the strength comes from the name next to it. This is a change in approach: before, version and power moved together (GPT-4, then GPT-5), now the generation and the capability level move on separate rhythms, so that Sol, Terra and Luna can each update on their own, without waiting for the next big version of the whole family.

In GPT-5.6, Sol is the top - the flagship model, the most expensive to use and the only one with access to the full "maximum mode" of thinking, a mode for the hardest multi-step tasks. Per unit of text you pay several times more through it than through the cheapest of the three, which says enough about the gap in ambition between them.

Between the two stands Terra - balanced for everyday work, made to do almost the same as the company's previous flagship, but at half the price. This is probably the model most products will actually use, because the trade-off pays off and the difference rarely shows.

Luna is the base - fastest, cheapest, for tasks that don't need deep reasoning, just a quick, good-enough answer. The volume runs through it: chatbots, small automations, everything that happens thousands of times a day and can't pay flagship price every time.

Codex is a different beast. It sits outside the Sol/Terra/Luna trio - OpenAI's separate coding agent, a system that goes into a project on its own, runs commands, fixes bugs and proposes code for review, instead of just answering a question. Since early July it's built into the ChatGPT app itself, and it launched on GPT-5.6 underneath. Over five million people use it weekly - a number that shows writing code is stopping being something you do alone in front of a screen. It's becoming something you hand off.

GPT-6: the answer becomes an interface

In September OpenAI released the first models of the GPT-6 generation, paid users first: Astra for the heaviest work, with Sol and Luna below it. On October 7, 2026 GPT-6 also came to the regular chat in ChatGPT, which by OpenAI's numbers more than 1.2 billion people use every week. The paid Plus, Pro, Business and Enterprise plans get GPT-6 Sol. The free plan and Go get GPT-6 Luna, and by OpenAI's plan the rollout to them starts the next day. The models in Work and Codex don't change with this release.

The new part you can see with the naked eye is called Intelligent UI. The model assembles the answer from text, pictures, buttons, forms, charts and small tools: a savings calculator, a bill splitter after dinner, a game inside the conversation itself. The interface appears piece by piece while the model is still generating it, and when plain text is the most useful answer, you get text. OpenAI itself admits the model's design judgment still has a way to go. The logic of the tiers stays: everyone gets GPT-6, just not the same one.

On October 26, 2026 Terra came back: OpenAI released GPT-6.1 Terra, a fast, efficient model for visual reasoning and precise computer use. It costs 1 dollar per million input tokens and 5 for output, and the context is 1,050,000 tokens. It takes text and images and returns only text. According to Vercel, OpenAI still recommends its larger GPT-6 models for the most complex coding. The tier for everyday work is the cheap one here too.

Splitting one model into three isn't a technical detail. It's an admission that intelligence is now a commodity with different price tags, not one miracle you get all at once.

Let's not dance around it

This split is something I've been doing at the agency for years: not every task deserves the most expensive person in the room. OpenAI just said out loud what the industry was already doing quietly. Sol, Terra and Luna aren't three different products under one label. They're one family with different characters - one is for the hardest problem, the middle one does most of the work, the small one runs fast and doesn't ask questions.

The real news for me isn't the trio of models. It's Codex inside them - a coding agent built into a model that decides on its own how much "to think" before writing a line of code. That's the direction everything I do here is heading. I no longer ask which model is "the best." I ask which one is good enough for this task, at this price, right now.

The visual is generated code-art, no third-party images.