How to pick the right AI model, now that every app has five of them
ChatGPT's picker has five reasoning levels. Claude has Haiku, Sonnet and Opus. Gemini has Flash and Pro. None of the names tell you which to use. The job does.
I don't know which model to pick. A plain way to choose between the fast, standard and heavy models in ChatGPT, Claude and Gemini as of October 2026: match the model to the job, not the benchmark. With the current picker names and a two-question test.
There are only three kinds of model
Ignore the names for a moment. Every company sells the same three things. A fast model that answers in a second and is right about simple things. A standard model that is right about most things and is the one you should live in. And a heavy model that thinks for a minute, costs ten times more, and is worth it for maybe one task in twenty.
As of late October 2026 the names are: in ChatGPT, a reasoning picker with Instant, Medium, High, Extra High and Pro, running on GPT-6 for paid plans and GPT-5.6 Luna on Free and Go. In Claude, Haiku 5.5, Sonnet 5.5 and Opus 5.5, with Fable 5.1 on plans that include it. In Gemini, Flash and Pro tiers. The names will change again by spring. The three kinds will not.
Ask two questions about the job
Before you pick, ask: does a wrong answer cost me anything? and does the task need more than three steps of thinking? If both are no, use the fast model and move on. If either is yes, use the standard model. Use the heavy model only when both are yes and you will read the whole answer.
The mistake I see most is the opposite one. People use the heaviest model for everything, wait a minute for each reply, burn through their weekly limit by Wednesday, and get answers that are longer but not better. A standard model with a clear prompt beats a heavy model with a vague one.
Set the default so you stop choosing
Decision fatigue is why people pick the wrong model. Set the standard model as your default and only change it for a specific chat.
In ChatGPT, the picker remembers your last choice for new chats; pick Medium once. In Claude, the model dropdown at the bottom of the chat box remembers your choice; set Sonnet. In Gemini, the model menu at the top of a new chat; set Pro if your plan has it, Flash if not. On a Free plan you mostly get the fast or standard model anyway, and that is fine for nearly everything.
Learn the one time to pay for heavy
The heavy model earns its cost in exactly one situation: a problem where the standard model has already failed and you can describe why. Paste the standard model's failed attempt, say what was wrong with it, and ask the heavy model to fix it. That gives the expensive model something to reason about instead of a cold start.
Do the maths on your own limit. On Claude, Opus uses the weekly budget several times faster than Sonnet. On ChatGPT, Extra High and Pro answers take a minute or more each. If you would not wait a minute for the answer, you do not need the heavy model.
A faster model tried this and got it wrong. Here is the task, its answer, and what was wrong with the answer. Task: [the task] Its answer: [paste] What was wrong: [one or two lines] Think it through step by step, show the reasoning, and give me a corrected answer. Tell me if you are unsure about any step.
You're done when
Questions people are asking
Which ChatGPT model should I use? Medium for most things. Instant for rewrites and quick questions. High or Extra High only for multi-step problems you will act on. Pro is for Pro-plan users with a hard problem and time to wait.
Is Claude Opus better than Sonnet? On hard problems, yes. On everyday work, Sonnet 5.5 is as good and several times cheaper against your usage limit. Use Opus when Sonnet has failed twice.
Do the model names keep changing? Yes. OpenAI retired its GPT-5.2 models in June 2026 and renamed its reasoning levels the same month. Pick by the three kinds, fast, standard and heavy, and the renames stop mattering.
Sources
- OpenAI Help Center, GPT-6 and other models in ChatGPT help.openai.com/en/articles/20001354-gpt-6-and-other-models-in-chatgpt
- OpenAI Help Center, ChatGPT release notes help.openai.com/en/articles/6825453-chatgpt-release-notes
- Anthropic, news and model announcements anthropic.com/news
- Google Gemini, release notes gemini.google/release-notes
How to stop ChatGPT and Claude losing the plot halfway through a long chat
One email a week. Written by people.
New how-tos, the tools we tested and what we think, in one short email.