Which ChatGPT model should I use? Instant, High or Pro, by plan and by job

Which ChatGPT model should I use for each job?

Instant for everyday questions, High for anything you will act on, and Pro only if you are on Pro $100 or above. Since 14 September 2026 that choice is yours alone on Plus and Pro: OpenAI retired the automatic switch from Instant to a thinking level, so a hard question typed into Instant stays on Instant. We read every OpenAI help page that describes the picker on 27 September 2026, and the rules below follow from what those pages say each option is for.

The picker has changed three times since June. On 10 June 2026 OpenAI collapsed Thinking Standard, Extended and Heavy into Medium, High and Extra High and dropped Thinking Light. On 6 August 2026 GPT-5.6 Sol took over every paid position with a thinking slider, and GPT-5.6 Luna became the Free and Go default. On 14 September 2026 the auto-switch went. If a guide you are reading still says "Thinking" or "Auto", it predates all three, and o3, GPT-4o and GPT-4.5 are gone from ChatGPT too (retired 26 August, 13 February and 27 June 2026).

What is in the ChatGPT model picker on each plan?

Six positions at most, on two models, and Plus sees three of them. Everything in the table is from OpenAI's GPT-5.6 and GPT-6 Pro in ChatGPT page and its Business and Pro tiers pages, read 27 September 2026. Personal and Business plans only; Enterprise workspaces can hide any row and use Pro Standard and Pro Extended labels instead.

Two lines on that page matter more than the table. The first: "On Plus and Pro plans, Instant doesn't automatically switch to a higher thinking level for complex requests." Asking it to "think harder" does not move it either, so the level you leave the picker on is the level you get. The second: "Terra and Luna are not selectable in standard ChatGPT conversations." The cheap GPT-5.6 family members exist, but only in Work, Codex and the API, so a paid chat is Sol or nothing.

Business is the odd plan out. Its picker carries the same six positions, but a Higher intelligence switch under Settings > General lets Instant add reasoning on its own, and OpenAI warns that "Automatic reasoning may use your reasoning allowance even while Instant remains selected." OpenAI also notes that "Pro is a model option, not another name for a Premium seat", which is the question its own support team evidently gets. Pro $200 has been closed to new sign-ups since 10 September 2026, so for anyone buying today the Pro row means Pro $100 and its 50 messages a week; ChatGPT Plus vs Pro works through whether that tier is worth it at all.

Which ChatGPT model should I use for coding?

High, in a chat, for a function or a bug you can paste; Codex for anything that spans files. OpenAI describes GPT-5.6 Sol as built for "coding, knowledge work and research, cybersecurity, science, computer use, and design", and Medium and High are the two Sol levels a Plus subscriber has, so High is the ceiling for pasted code on a $20 plan.

The stronger coding models are not in the chat picker at all. Codex and ChatGPT Work got GPT-6 Sol and GPT-6 Luna on 22 September 2026, and OpenAI says they "are not selectable in regular ChatGPT conversations". GPT-6 Astra is in Codex for Plus and above, with Business Standard seats estimated at 5 to 45 Astra messages per five-hour window against 10 to 100 for GPT-5.6 Sol and 250 to 2,000 for Luna. So the honest coding rule on Plus is: chat on High until the task touches a second file, then open Codex, where the plan's better models live and where the allowance is separate from chat.

Which ChatGPT model should I use for image generation and deep research?

Neither is a model choice. Image generation runs through the Images tool (ChatGPT Images 2.5 since September 2026) whatever position the picker is on, and deep research is a task you start from the + menu, by typing /Deepresearch, or from the sidebar. The picker decides how the text model thinks; these two ride alongside it.

What your plan changes is the allowance. OpenAI's pricing page gives Free "Limited and slower image generation" and "Limited deep research", Go "More image creation", Plus "More complex and accurate image creation" and "Expanded deep research", and Pro "Unlimited and faster image creation" and "Maximum deep research". OpenAI no longer prints a task count for deep research; its help page says "Your in-product usage counter shows your remaining tasks" and that a fixed monthly allowance "resets every 30 days from the date of your first use". Deep research "is powered by the latest models" by default and you can pick a legacy model for a run if you prefer, which is the one place the model choice still touches it. Google's equivalent is more explicit: Gemini Deep Research reports take "about 5-10 minutes", all users get the Thinking model for them, and AI Pro and AI Ultra subscribers can run them on Pro.

What is the equivalent choice in Claude and Gemini?

Claude asks you to pick a model and an effort level; Gemini asks you to pick a model and a thinking level; ChatGPT on a paid plan asks only for the level, because the model is fixed. Same three vendors' help pages, read 27 September 2026, consumer apps only.

The last row is the market fact this page exists to state. ChatGPT is the only one of the three whose free tier offers no model choice: Anthropic's guide says "Free includes Haiku and Sonnet", and Google's limits page lists all three Gemini models as available "Without an AI Plan", with the plan changing the context window (32k free, 128k on AI Plus, 1 million on AI Pro and Ultra) and the usage multiple (2x, 4x, then 5x or 20x) rather than the models. OpenAI's free tier gives you Luna and a Think button, and the choice begins at $20.

Anthropic's math advice matches the ChatGPT rule above. Its settings page says to raise effort, turn on thinking, or both for "Mathematical calculations and proofs" and "Competition-level coding challenges", and that Extra high effort "is designed for long-running coding and agentic tasks". Google says its Pro model "excels at complex math and coding prompts". Three vendors, one pattern: the fast position for questions, the extended position for anything you will check. Best AI model for every task picks between the vendors per job; this page is about the picker inside one of them.

Should you let a router pick the model for you?

On ChatGPT Plus and Pro you no longer can, and on the other apps the router is a default rather than a switch. OpenAI's June 2026 picker still let Instant "auto-switch to Medium for higher reasoning when required"; the September change removed that for personal plans and kept it only in Business, behind the Higher intelligence setting. Claude routes by giving each model a recommended effort marked "Default" and leaving the model to you. Gemini gives every plan all three models and lets usage limits do the steering: when you hit one, "you can continue your conversation with Flash-Lite".

Whizi, which publishes this page, runs one more router. Its Auto row sorts each message onto one of six fixed rungs across five models from three vendors, and a turn costs whatever the picked model costs: 1 credit on Starter, 1 to 20 on Pro and Powerhouse (trial week from $0.99 on Starter). Where it loses to the ChatGPT picker: it routes across vendors, so it cannot read your ChatGPT memory, Projects or custom GPTs; on Starter it never leaves the house model; and GPT-6 Astra is a hand-pick on Powerhouse at 100 credits a message, not a rung Auto reaches.

The decision rule, then, for a ChatGPT subscriber in September 2026: leave the picker on Instant, move it to High before any question whose answer you will paste somewhere, and only think about Pro if you are already paying $100 and have 50 weekly messages to spend. Everything else the picker offers is a level of the same model. How to choose an AI model is the ten-minute test for deciding whether that model should be GPT at all.