Between early September and 26 October 2026, OpenAI replaced almost every model inside ChatGPT. There is no longer one "GPT-6". There are four, named Luna, Terra, Sol and Astra, and which one answers you depends on your plan and on which part of ChatGPT you are using. That sounds like a detail, but it explains most of the "ChatGPT got better" and "ChatGPT got worse" posts you have seen this month. Two people can type the same prompt and get answers from very different models.
Four models, one family
OpenAI now sorts its models into four classes, ordered from cheapest and fastest to most capable. Luna is the small, quick model for focused, high-volume work. Terra is the medium model. Sol is built for complex coding and agentic work, meaning tasks where the AI takes several steps on its own. Astra sits at the top. The first GPT-6 model, Astra, arrived in early September. Sol and Luna followed on 22 September, when OpenAI said both were 50% cheaper than the GPT-5.6 versions they replaced.
The newest member is GPT-6.1 Terra, released on 26 October. OpenAI describes it as a fast, efficient model that combines detailed visual reasoning with precise computer use. In practice, that covers jobs like reading a technical drawing, filling in a web form or drafting a proposal from a screenshot. It is available in ChatGPT Work, Codex and the API, at about half the price of the GPT-5.6 Terra it replaces.
Which model answers you
On 7 October, OpenAI began rolling GPT-6 Sol out in regular ChatGPT for Plus, Pro, Business and Enterprise subscribers, and GPT-6 Luna to Free and Go users. That is the single most important fact in this story. If you are on the free plan, you are using the smallest model in the family. It is a good model, but when a free user and a paying user compare answers to a hard question, they are not comparing like with like.
Where you use ChatGPT matters too. ChatGPT Work, the agent mode that carries out longer tasks across apps, and Codex, the coding agent, got the new models first, and they are the only places GPT-6.1 Terra shows up so far. If you do most of your work in the plain chat window, the Terra upgrade won't reach you yet.
Error rates at low reasoning effort on OpenAI's internal set of error-inducing prompts, as reported by 9to5Mac on 26 October 2026. Lower is better.
That drop is the most practical improvement in this release. OpenAI says the error rate on its set of deliberately tricky prompts fell from 53.6% to 16.7% at low reasoning effort, and to 4.3% at maximum effort. These are OpenAI's own internal numbers, so treat them as a direction rather than a guarantee, but they match what people expect from a newer model: fewer confident wrong answers on the fast setting.
Intelligent UI changes what an answer looks like
The other big change arrived when GPT-6 came to ChatGPT on 7 October. OpenAI calls it Intelligent UI, and it means ChatGPT no longer has to answer in paragraphs. A response can now include charts, tappable buttons, forms and small interactive tools that work inside the conversation. OpenAI's examples are a side-by-side layout when you compare things, an interactive diagram when you ask for an explanation, and plain text when that is clearest. ChatGPT can also build little tools on request, like a calculator or a bill splitter.
The skill to learn here is asking for the format you want. "Compare these three laptops" might produce a table. "Build me a tool where I can change the budget and see which laptop wins" produces something you can play with. If you work with numbers, try asking for an interactive version of any calculation you redo every month, such as a pricing scenario or a hiring budget, and save the chat so you can come back to it.
What we would do this week
First, check what you are actually using. Open the model picker in ChatGPT and see which GPT-6 model is selected. If you are on a paid plan and your answers feel thin, make sure you haven't been switched to a lighter model, and choose the longer thinking option for anything important if your plan offers one.
Second, if your work involves screenshots, forms or documents with diagrams, try the same task in ChatGPT Work with GPT-6.1 Terra. Visual reasoning and computer use are exactly where OpenAI says Terra improved most, so that is where you will notice the difference.
Third, don't rush to rebuild your setup around ChatGPT if you mainly use another assistant. Claude and Gemini both shipped new models in the same weeks, and every lab is now on a roughly monthly release rhythm. Choose a tool based on the jobs you do every week, not on the latest benchmark chart. Our Compare page sets ChatGPT, Claude and Gemini side by side for common tasks, and our ChatGPT lessons walk through Projects and the other features most people never set up.
One more deadline is worth knowing about. OpenAI retires custom GPTs on 11 December 2026 and is moving them to plugins. If you or your team rely on a custom GPT, plan its migration now rather than in December.