Models: GPT-6 and Thinking Levels

Understand the models available in ChatGPT

When you open ChatGPT, you aren’t always talking to the same model. In fall 2026, the GPT-6 family replaced the GPT-5.6 family from summer. Choosing the right setting is a skill in itself: it’s the difference between waiting a minute for an answer that should have taken three seconds, and entrusting a delicate arbitration to the fastest setting.

First, a distinction that clears up confusion: Chat and Work don’t have the same menu. In Chat, you choose GPT-6 and a thinking level. In Work and Codex, you choose among four models, with a power slider. Depending on your plan, workspace, and current deployment, you’ll see all or some of the options described here.

In Chat: GPT-6 and the thinking slider

In the Chat model menu, GPT-6 replaced the old “Latest” choice. A Thinking slider then sets the thinking effort:

• Instant: a fast response, for everyday questions;
• Medium: standard thinking;
• High: deep thinking;
• Extra High: the highest level of GPT-6 Sol;
• Pro: the Pro models, for difficult tasks and long work; on eligible plans, this is GPT-6 Pro, powered by GPT-6 Astra.

On Plus, Pro, Business, and Enterprise, GPT-6 Sol responds from Instant to Extra High. On Free and Go, ChatGPT relies on a lighter Luna model and offers a Think button for harder questions. Accessible levels depend on the plan:

Three useful clarifications. Older models remain selectable: if you explicitly choose GPT-5.6 Sol, ChatGPT won’t switch you to GPT-6. If you hit a thinking level limit, ChatGPT may continue with another available model and indicates the renewal time when possible. Finally, Intelligent UI enriched responses (visuals, interactive elements) work from Instant to Extra High, not with Pro.

In Work and Codex: four models

Work and Codex have their own models. GPT-6.1 Sol and GPT-6.1 Terra aren’t offered in Chat, and Astra only appears there as the Pro level.

• GPT-6 Astra, the most capable: for the most demanding work, requiring reasoning, research, and end-to-end document production.
• GPT-6.1 Sol, almost Astra, cheaper: near-Astra performance for a lower cost, a great starting point for demanding daily work.
• GPT-6.1 Terra, fast and efficient: for clear writing, image understanding, and precise control of a browser or computer.
• GPT-6 Luna, light and economical: for targeted, repetitive tasks; this is the Work and Codex model on Free and Go plans, in the desktop app.

The principle is the same across the range: there are no more separate “reasoning” models like o3 or o4-mini. Each model can answer fast or think long, and you set that intensity. Recent release dates: GPT-6 Astra on September 3, GPT-6 Sol and Luna in Work and Codex on September 22, GPT-6.1 Sol on September 29, GPT-6 in Chat on October 7, and GPT-6.1 Terra on October 26, 2026.

The power setting

In the desktop app and in ChatGPT Work on the web, OpenAI recommends starting with the default Power setting for your account, then sliding to Smarter for deeper reasoning, or to Faster for quicker, cheaper work. Each step combines a model and a reasoning level, for example “Luna High”, “GPT-6 Sol Light”, or “Astra Medium”. The Advanced panel lets you choose precisely a model, reasoning effort, and speed. Options and default values vary by plan, app, workspace settings, and deployment: don’t be surprised if you don’t see exactly the same steps as a colleague.

Two extreme settings are worth knowing. Max gives the model more time to reason on a single task: reserve it for the hardest problems, when depth matters more than speed. Ultra goes further by delegating parts of a complex task to sub-agents working in parallel; it’s only useful if the work splits into independent chunks. GPT-6 Luna goes up to Max, but doesn’t offer Ultra. As OpenAI reminds us, most tasks need neither.

How to choose, and the most costly mistake

In Chat, stay on the default setting for common questions, and move up one step when the question requires weighing arguments, chaining steps, or verifying reasoning. In Work, a targeted, repetitive task calls for Luna; text to write, an image to understand, or an action in an app, Terra; demanding daily work, GPT-6.1 Sol; a truly complex case, Astra.

The most common mistake is pushing the slider to the maximum to rephrase a sentence: you wait, you consume your quota, and the result isn’t better. The symmetrical mistake is more discreet: never touching the setting and entrusting contract analysis to the fastest step. Glance at the active model and level before an important task.

Your turn

Compare for yourself. Ask exactly the same question twice, first with the default setting, then with a higher thinking level (or the Think button, on Free and Go):

“What are the pros and cons of creating a SAS or a SARL for a 3-person tech startup in France? Present the key differences in a comparison table.”

Observe the response speed, then the analysis depth. Keep in mind that a longer answer isn’t necessarily more correct: on a legal question, always check decisive points with an official source like service-public.fr.

Official sources: GPT-6 and other models in ChatGPT (OpenAI Help Center: thinking levels, plans, limits) · ChatGPT Work and Codex (Work and Codex models) · Models (ChatGPT documentation: power setting, Max and Ultra) · ChatGPT release notes (release dates)

Key takeaways

• Chat and Work don’t have the same menu: GPT-6 and a thinking level on one side, four models and a power slider on the other
• In Chat, the slider goes from Instant to Extra High, then Pro; Free and Go have a Luna model and the Think button
• In Work and Codex: Astra, GPT-6.1 Sol, GPT-6.1 Terra, and Luna, with Smarter, Faster, Max, and Ultra settings
• There are no more separate reasoning models: you set the thinking intensity for each model
• Adapt the setting to the task: neither maximum for a triviality, nor minimum for an important decision