Which ChatGPT Model Should You Use for What?

TL;DR: Which ChatGPT model to use for what, verified against OpenAI's own Help Center on September 5, 2026: paid plans default to GPT-5.6 Sol across Instant, Medium, High, and Extra High reasoning; Free and Go default to GPT-5.6 Luna; and GPT-6 Astra is rolling out for complex computer-use and coding work, first to a limited group and then to Plus, Pro, Business, and Enterprise.

What Does the ChatGPT Model Picker Actually Show You Right Now?

Here's the part that trips people up: the picker you see inside ChatGPT today does not show you raw model names. It shows you reasoning-effort levels, and OpenAI decides which underlying model answers each level based on your plan. As of September 5, 2026, per OpenAI's own Help Center article on GPT-5.6 in ChatGPT, the options are Instant, Medium, High, Extra High, and Pro, with a dedicated reasoning slider that lets eligible paid plans choose how much the model thinks before it answers.

Underneath that slider, on Plus, Pro, Business, and Enterprise, all four non-Pro levels run on one model, GPT-5.6 Sol, and only differ in how long it's allowed to reason. The Pro level is a separate, higher-capability option that includes GPT-5.6 Sol Pro and, as it continues rolling out, GPT-6 Pro. On Free and Go, the picker looks different: those plans don't get Sol at all. Their default is GPT-5.6 Luna, and the same model also powers the Think option those plans get for harder questions.

This is a change from how ChatGPT used to work. Earlier in 2026, the picker showed named model-and-mode pairs like GPT-5.5 Instant and GPT-5.5 Thinking that users chose directly. The reasoning-slider interface is the newer layer on top of the GPT-5.6 family, and OpenAI's Help Center notes that enterprise workspaces "may continue to see the existing model picker until the new controls reach their workspace" — so what you see can genuinely differ from a colleague's screen for a few weeks during any given rollout.

Which ChatGPT Model Should You Use for Which Task?

None of this replaces writing a clear prompt. Picking the right reasoning level and then asking a vague question still gets you a vague answer — the model choice sets a ceiling, the prompt decides whether you reach it.

Which Model Do Free and Go Users Actually Get?

Free and Go users are on GPT-5.6 Luna by default, full stop — OpenAI's Help Center is explicit that these plans "do not have access to GPT-5.6 Sol." Luna is described in the same documentation as the fastest and lowest-cost model in the GPT-5.6 family, which is exactly why it's the plan tier that gets the lightest-weight one. When a Free or Go question needs more depth, the Think option is the only lever available, and it still runs on Luna rather than switching you to Sol — a distinction worth knowing before you assume that thinking harder on a free plan gets you the same reasoning quality as a Plus subscription.

One thing that has not changed with any of this: OpenAI's pricing page states that Free and Go users get unlimited everyday text chats, subject to its abuse-prevention safeguards, with separate and tighter limits on file uploads, image generation, voice, and data analysis. If your work is mostly plain conversation, the free tier's ceiling is higher than most people assume; it's the auxiliary tools that get capped first. Go sits just above Free: OpenAI's own Help Center describes it as a low-cost plan giving "expanded access to ChatGPT’s most popular features". It raises the ceiling on image generation, file uploads, and data analysis without adding GPT-5.6 Sol — the model you're on stays Luna either way, and what changes between Free and Go is how much of everything else you get around it.

How Do Business, Enterprise, and the Reasoning Levels Actually Line Up by Plan?

The reasoning-slider levels aren't uniformly available even across paid plans. OpenAI's own comparison table, current as of September 5, 2026, breaks it down like this:

The gap worth noticing: Plus gets Medium and High, but not Extra High or the Pro tier. If your work regularly needs the hardest reasoning setting, Plus doesn't unlock it, only Pro, Business, and Enterprise do. Business and Enterprise workspaces add one more variable on top: an admin can restrict which models workspace members can use at all, so a corporate account might show a narrower picker than an individual subscription at the same nominal plan level, and it's worth checking with whoever administers your workspace before assuming the table above applies unmodified to your seat.

Does the Model You Use for Chat Match What Codex and Work Use?

No, and this is where a lot of confusion starts for anyone who also codes with ChatGPT. The chat model picker described above is a separate surface from ChatGPT Work and from Codex, OpenAI's coding-agent product, and availability doesn't mirror one-to-one. Per OpenAI's Help Center: in Work, Plus, Pro, Business, and Enterprise get all three GPT-5.6 models, Sol, Terra, and Luna; in Codex, Free and Go get Terra only, while Plus, Pro, Business, and Enterprise again get all three. That's the one place Terra becomes directly selectable rather than something only running behind the scenes.

If you're pairing ChatGPT with actual coding work, two version numbers are worth checking before you assume you have current model access there at all: OpenAI's article on GPT-5.6 lists a minimum ChatGPT desktop app build of 26.707.30751 for Codex mode, and a minimum Codex CLI version of 0.144.0. An outdated client is a separate reason you might not see a model the picker above says you should have, distinct from a plan or workspace restriction.

When Should You Reach for Medium, High, or Extra High Instead of Instant?

The honest answer is: less often than the picker's four options might suggest. Instant on GPT-5.6 Sol isn't a stripped-down mode, it's the same model family with a shorter reasoning budget, and OpenAI's Help Center describes an automatic-reasoning setting, Higher Intelligence, that lets Instant quietly think longer on a complex request without you manually reaching for a slider at all. For most day-to-day questions, that automatic behavior plus Instant covers the job.

Reach for Medium or High deliberately when a task has several steps that depend on each other: comparing claims across multiple sources, restructuring a document while preserving specific constraints, or working through a coding problem where an early wrong assumption breaks everything downstream. Extra High is worth the extra wait mainly for the hardest end of that range — the kind of task where you would, if delegating to a person, explicitly ask them to double-check their own work before sending it back. If you consistently need to ask a lot of follow-up questions to get a full answer, that's also a signal to move up a reasoning level rather than rephrasing the same prompt five times.

What Is GPT-6 Astra, and Should You Be Using It Today?

This is the fastest-moving fact on this page, which is exactly why it needs a date stamped on it rather than a superlative. OpenAI introduced GPT-6 Astra on September 3, 2026, positioning it for computer use, browsing, software engineering, cybersecurity, science, and professional document and spreadsheet work. Its own release page states the rollout began with a limited set of organizations in OpenAI's Daybreak Access Program, with availability for ChatGPT Plus, Pro, Business, and Enterprise users, and the OpenAI API, following within days of that announcement.

If your model picker doesn't show it, that's expected, not a bug: reasoning-effort levels on GPT-5.6 Sol remain the correct choice for the task table above, and nothing in the picker degraded while Astra rolls out around it.

What Happened to GPT-4o, GPT-5, and Older Models?

If a search result or an old bookmark tells you to pick GPT-4o or GPT-5 Thinking, that advice is out of date, and knowing when helps you judge how stale the rest of the page probably is. Per OpenAI's own Model Release Notes, retirements from ChatGPT (not the API, which followed its own separate schedule) landed on these dates:

In each case, OpenAI rolled existing conversations forward onto the nearest current equivalent rather than simply breaking them, and paid plans retain a "Legacy models" toggle in settings for models that haven't been fully retired yet. That toggle is worth checking before you assume a model you remember by name is actually gone — but treat any list you find of what's currently inside it as perishable, since it changes with every retirement announcement above.

Do You Even Need to Pick a Model Manually?

Less than the existence of a picker implies. The single highest-leverage setting most people skip is Higher Intelligence, in Settings > General, which lets Instant reason longer automatically on paid plans without you touching the slider at all. Manual selection earns its keep in two situations: when you already know a task is going to be long and multi-step before you start, so you'd rather not wait for automatic detection, and when you're deliberately trying to compare how the same prompt performs at two different reasoning levels.

That second use case is worth doing at least once. Run an identical, moderately complex prompt through Instant and then through High, and read both answers side by side. It's the fastest way to develop an intuition for what "more reasoning" actually buys you on your own work, rather than trusting a generic table like the one above for every task you'll ever hit. And if either answer comes back thin regardless of the level you picked, the fix is usually the prompt, not the model — a one-line request produces a one-line-shaped answer at any reasoning effort.

How Do You Keep This Page From Going Stale on You?

Everything above is dated on purpose, because it will not stay current. OpenAI has retired, renamed, or meaningfully changed a ChatGPT model roughly every four to eight weeks across 2026 according to its own release notes, and the GPT-6 Astra rollout described above, still in progress as this is published on September 3-5, 2026, is one more instance of that same pattern. Set an actual recurring calendar reminder, monthly is reasonable, to reopen OpenAI's Help Center model article and glance at your own picker. Six months of confidence in a name you memorized once is exactly how just use GPT-4o advice keeps circulating on forums a full retirement cycle after it stopped being true.

If you're weighing ChatGPT's models against Claude's instead of against each other, that's a different, equally dated question, covered separately in ChatGPT vs Claude: which writes better prompts. And whichever model you land on, the same underlying discipline applies regardless of which one is current when you read this: a specific, well-scoped prompt outperforms a vague one at every reasoning level, which is the whole premise behind 100+ ready ChatGPT prompt templates built to be specific by default.

Save this as your own quick reference, and swap the model names for whatever OpenAI is shipping the day you read it:

CHATGPT MODEL QUICK REFERENCE (verify against help.openai.com before trusting)

Everyday questions, casual writing .......... Instant (auto-boosted by
                                                Higher Intelligence when needed)
Multi-step research, source comparison ....... Medium / High
Code with real dependencies, debugging ....... High / Extra High
One hard, high-stakes deliverable ............ Pro tier (Sol Pro / GPT-6 Pro)
Free or Go plan, harder question ............. Think (still runs on Luna)
Computer-use / multi-app agentic work ........ GPT-6 Astra, IF it has reached
                                                your plan yet — check the
                                                picker, don't assume

Recheck monthly: help.openai.com > ChatGPT > "GPT-5.6 [or current] in ChatGPT"