Choosing Models and Reasoning Effort in ChatGPT: GPT-6 Sol, Astra, Luna and Terra

Version note: This post describes the model and reasoning effort options in ChatGPT as of 15 November 2026; the model menu, plan differences, usage estimates and retirement dates were checked the same day on the Models and Pricing pages of the ChatGPT documentation (learn.chatgpt.com) and in the Help Center's model article. This post is based on the official documentation. The model menu changes with the plan, the surface and gradual rollout, so the options on your screen may differ.

Guess First

Ayşe is a developer on Bookstore, the online bookstore used as the shared example across this site's series, and she is on Plus. On Monday she noticed three things. First, she wanted to follow a blog's advice to "set Thinking to Extra High in Chat", but Extra High was not in her menu. Second, when she switched to Work she saw a slider labelled "Smarter" and "Faster" instead of model names. Third, a weekly scheduled task she had set up in September, which selects gpt-5.5 in Codex, had been failing since mid-October.

What is your guess? Which of these three is a bug, and which is normal behavior of her plan or the product? And is it better for Ayşe to do every job at the highest setting, or to choose per job? The answer is in the Help Center's model article and the Models page on learn.chatgpt.com.

The Model Menu and Thinking Setting in Chat

The GPT prefix in OpenAI's model names reached the GPT-6 family in 2026. According to the Help Center article "GPT-6 and other models in ChatGPT" (https://help.openai.com/en/articles/20001354, 15 November 2026), GPT-6 replaced the option that used to appear as "Latest" in Chat's model menu. With GPT-6 selected, a Thinking setting chooses how much thought the model puts into an answer: every level from Instant to Extra High runs on GPT-6 Sol. The Pro option in the menu uses GPT-6 Pro, which runs on GPT-6 Astra behind the scenes. According to the release notes, GPT-6 started rolling out in Chat on 7 October 2026 for Plus, Pro, Business and Enterprise.

The plan differences answer the first question from Guess First. Plus has Medium and High but not Extra High or Pro; Extra High missing from Ayşe's menu is not a bug but a plan limit. Free and Go have no Medium, High or Extra High at all; their default model comes from the Luna family. The sources disagree here: a September 2026 announcement gives GPT-5.6 Luna as the default for Free and Go, while the 7 October GPT-6 note says GPT-6 Luna for those plans. That is a good example of this post's main advice: do not memorize model names, check your own menu.

Two more details are worth knowing. First, GPT-6 Pro has its own usage limits; a Pro subscription does not mean unlimited Pro use, and the limits differ between the 100-dollar and 200-dollar Pro tiers. Second, this setting in Chat only affects Chat. The September 2026 announcement says explicitly that the Thinking slider and model changes in Chat do not change model behavior in Work or Codex. What you pick in Chat does not decide which model runs in Work.

In Work and Codex: Power Presets and the Advanced Menu

In Work and Codex the choice is in the model and reasoning control below the composer. According to the Models page on learn.chatgpt.com (https://learn.chatgpt.com/docs/models, 15 November 2026), in the desktop app and in Work on the web you start with your account's default Power setting; you move the slider toward Smarter for deeper reasoning or toward Faster for faster work that uses less. The documentation's example slider has six presets: Luna High, GPT-6 Sol Light, GPT-6 Sol Medium, Astra Light, Astra Medium and Astra Extra High; some paid plans do not show Astra Extra High. To choose a specific model, effort level or speed, you open the Advanced menu; models that are not on the slider, such as GPT-6.1 Terra, are picked there.

That explains Ayşe's second observation: the "Smarter/Faster" slider in Work is a simplification, and the model names sit under Advanced. The documentation also notes that the options and defaults shown vary by plan, client, workspace settings and rollout. In Enterprise and Edu, GPT-6.1 Sol is off by default and an admin has to turn it on; Astra needs admin approval as well.

The Model Family (15 November 2026)

These are the recommended models in Work and Codex, with two indicators from the Pricing page. The five-hour estimate is the number of local messages in one five-hour window on Plus and Business standard; the credit rate is the number of ChatGPT credits used per 1 million input and output tokens at Standard speed. Credits are the unit for paid extra usage, but they also show how "heavy" the models are relative to each other.

The gap is large. For the same credits, Luna processes about a hundred times as many tokens as Astra; in the five-hour estimate, Luna's upper bound is more than sixty times Astra's. The release history is fast too: Astra arrived in the week of 31 August, GPT-6 Sol and Luna in the week of 21 September, GPT-6.1 Sol in the week of 28 September and GPT-6.1 Terra in the week of 26 October. GPT-6.1 Sol's default context window in Work and Codex is 500,000 tokens.

Reasoning Effort, Max and Ultra

The Models page defines reasoning effort as how much work the model puts into thinking through a request. Higher effort can improve results on complex tasks, but it takes longer and uses more tokens. In the desktop app and in Work on the web the levels are Light, Medium, High and Extra High (the lowest one is called Low in the CLI). The documentation recommends Light for quick, well-scoped tasks, Medium for tasks that need some planning, and High and Extra High for difficult work with multiple steps, sources or trade-offs. As a starting point it gives the client's default for GPT-6.1 Sol, High for Luna and Light for Astra.

There are two more levels. Max gives the selected model more time to reason about a single task; it is meant for the hardest problems, where depth matters more than speed or usage. Ultra goes beyond a single agent and hands separate parts of the work to subagents in parallel; it only helps when the work really splits into meaningful parts. The documentation says plainly that most tasks need neither Max nor Ultra. GPT-6 Luna supports up to Max but not Ultra. If Ultra does not appear on the slider for a model that supports it, turn on "Ultra in model picker slider" under Settings > Configuration.

Separate from effort there is a speed setting. According to the Pricing page, Fast mode uses included usage 2.5 times faster than Standard, and Ultrafast mode on Astra and GPT-6.1 Sol 8 times faster; Ultrafast is only on the 500-dollar Pro tier and on eligible Enterprise and Edu plans. Effort levels also do not map exactly between model generations: High on GPT-5.6 is not the same as High on GPT-6. That is why the documentation suggests trying a familiar task at a lower setting and adjusting based on the result.

Job → Model Mapping

Translating the "Where each model shines" section of the Models page into the Bookstore team's work gives this starting point. It is not a rule; it is a first guess you will correct as you try your own work:

• Turning meeting notes into an action table, sorting 200 customer reviews into categories, extracting specific fields from a CSV: GPT-6 Luna, High. Repetitive work where you know what a good result looks like.
• A proposal text for institutional sales, interpreting a dashboard screenshot, filling a form from a source document: GPT-6.1 Terra, default effort.
• Combining a PRD, meeting notes and a timeline into an implementation plan, long-running and repeated document and code work: GPT-6.1 Sol, default effort, High if needed.
• Comparing three vendor proposals against requirements and recommending one, an end-to-end workflow across code, apps and research: GPT-6 Astra, start with Light.
• A short question in Chat, understanding an error message, fixing the tone of an e-mail: GPT-6, Instant.

The documentation stresses one more thing: clarify the task before you change the model. Astra's strength shows when you give it the sources, templates, constraints and checks; running the largest model on a vague request just buys a vague result at a higher price.

The Model Retirement Cycle

Ayşe's third observation is a retirement. According to the Models page, GPT-5.5 was removed from Codex with ChatGPT sign-in on 14 October 2026, while the "What's new" note for the week of 14–18 September, which gives the same date, describes the scope as ChatGPT, ChatGPT Work and Codex. The OpenAI API and Codex sessions signed in with an API key were not affected. The recommended replacement is GPT-6.1 Sol on Plus, Pro, Business, Enterprise and Edu, and GPT-6 Luna in the desktop app on Free and Go. Before that, GPT-5.4 and GPT-5.4 mini were removed on 31 August, and GPT-5.3-Codex-Spark on 14 September 2026.

The documentation also lists what needs updating: workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks and scripts that select a model. Ayşe's task is right on that list. On the Chat side, the Help Center says older models selectable from the menu are not being retired for now. The rule is simple: if you hard-code a model name somewhere, that place needs an owner and a review schedule.

Bookstore: Trying the Campaign Pricing Analysis at Two Settings

Selin wants to decide between two options for the November campaign: 20 percent off the whole fiction category, or 30 percent off 50 slow-selling backlist titles? She has an anonymized CSV of third-quarter order data and a document describing the pricing rules. The team started arguing about "should we run this at the highest setting?". The way to settle it is to give the same task at two settings and compare the results against the same criteria:

Attached: orders-2026-q3.csv (one row per order line: isbn, category,
list_price, discount_rate, quantity, order_date) and pricing-rules.md.

Question: for the November campaign, compare two options:
A) 20% off every book in the Fiction category
B) 30% off the 50 backlist titles with the lowest Q3 sales

Using only the attached data, estimate revenue and gross margin for both
options. Show each calculation step and the formulas you used.
List your assumptions separately. List what this data cannot tell us.
End with a recommendation and the one number that would change it.

First send it in Chat with GPT-6 at Instant, then send the same message in a new chat at High. If you have Work access, a third run can use GPT-6.1 Sol in Work at the default effort; this job is a file analysis, and Work's ability to run code for the calculations helps here. Based on the documentation's definitions, the expected difference goes in this direction: the low setting answers faster and uses less; the high setting does more planning and checking for the multi-step calculation and the trade-off between the two options, but takes longer and takes more of your limit. Whether either is good enough for your job can only be judged by comparing the results against the criteria below:

The point of this table is not to pick a winner but to let the team derive its own rule. If the Instant answer gets the calculation right and lists the assumptions, this kind of analysis does not need High; only if the trade-off row shows a clear difference does the high setting become the rule for that job. The LLM Fundamentals series post "LLM Cost, Latency and Model Choice: From Tokens to the Bill" explains the API-side logic of this comparison; here we apply the same logic through the product's menu.

In Production

Start with the default. The Models page and the web documentation say the same thing: start with the default effort and raise it when the task needs deeper planning or analysis. One person's habit of "do everything at Extra High" directly increases credit spending in a shared Business workspace.

If you pin a model name somewhere, own it. If a scheduled task, custom agent or script contains a model name, someone has to follow retirement announcements. The What's new page announced the last two retirements about a month in advance; the warning for GPT-5.5 came in the week of 14–18 September.

In Enterprise, the menu depends on the admin. If you cannot see Astra, GPT-6.1 Sol or Luna, ask your workspace admin first; the documentation says these models have to be enabled by an admin in Enterprise and Edu.

Common Pitfalls

The highest effort for every job. The documentation says plainly that most tasks need neither Max nor Ultra. Higher effort takes longer, uses more tokens and burns through the limit quickly.

Expecting every model on Free and Go. On these plans Chat has no Medium, High or Extra High; Work and Codex are limited and only in the desktop app. Check your plan before comparing a screen from a blog with your own.

Thinking the Chat setting affects Work. The Thinking setting in Chat and the Power preset in Work are separate controls; neither changes the other.

Giving Ultra a job that cannot be split. Ultra spreads the work across subagents in parallel; for a job that needs one sequential line of reasoning, that parallelism gains nothing and only adds usage.

Memorizing model names. Even the table in this post will be out of date within a few months. Know which model a saved setting or script selects, and follow the What's new page.

A Small Experiment

★☆☆ Open the model menu in Chat and in Work and write down the options you see. Compare them with the table in this post: which options does your plan hide?

★★☆ Adapt the Bookstore campaign prompt to your own data (for example, a cost comparison from your own team) and run it at two settings. Fill in the evaluation table; in which row did you see a real difference?

★★★ Write a "job → model and effort" rule page for your team: the five most frequent jobs, with a model, an effort level and a reason for each. List every scheduled task and script that names a model and assign an owner to each.

Whatever model you choose, ChatGPT also carries what it knows about you into every conversation. The third post covers personalizing ChatGPT with base style and tone, custom instructions, memories and Temporary Chat, and why required rules belong in a project rather than in memory.