Which AI Model Should You Use? A Simple Guide to Fast, Thinking and Pro Modes

Verified against vendor documentation on September 18, 2026. Model names, plan access and usage limits change frequently.

Open ChatGPT, Claude or Gemini today and one question can appear before you’ve even asked your actual question:

Which model should I use?

Then comes another one. How much should it think?

For experienced AI users, those menus are manageable. For everyone else, the obvious solution is to choose the most capable option available and leave it there.

That is usually overkill.

The useful rule is simpler:

Start with the normal model and its default reasoning level. Move up when the problem gives you a reason to.

Think of AI models as gears rather than a leaderboard. First gear is not a bad gear. Fifth gear is not a better gear. The right choice depends on what you’re asking the machine to do.

This is a narrower question than which assistant fits the way you work. The provider is already chosen. What remains is which of its models and reasoning levels to select for the task in front of you.

The four-level shortcut

If the answer is already accurate, useful and follows your instructions, there is usually little reason to make the model think longer.

If it starts losing constraints, making shallow assumptions, failing repeatedly or struggling to plan several steps ahead, that is your cue to change gear.

ChatGPT: the model choice matters more after September 14

OpenAI changed an important part of ChatGPT this week.

On September 14, OpenAI announced that it was retiring automatic switching from Instant to Thinking for ChatGPT Plus and Pro users globally. It also said it was removing the Higher intelligence setting from ChatGPT on the web for those plans. Automatic switching can still occur for safety purposes.

That means users need a clearer idea of what the reasoning selector actually does.

As of September 18:

OpenAI’s GPT-5.6 and GPT-6 Pro documentation says Instant provides fast responses for everyday questions, while Medium, High and Extra High progressively increase GPT-5.6 Sol’s thinking. Its Pro options include GPT-5.6 Sol Pro and, on eligible plans, GPT-6 Pro, which runs on GPT-6 Astra.

So if you are asking ChatGPT to improve an email, explain a concept or summarize something, Instant is a sensible starting point.

If you’re asking it to compare competing evidence, work through several constraints or solve a difficult technical problem, moving to Medium or High begins to make sense. Choosely covered the arrival of those reasoning controls in ChatGPT when GPT-5.6 Sol started rolling out.

A Plus subscriber does not need to hunt for Extra High or Pro. Those aren’t currently included on Plus.

That distinction matters because “use the strongest setting” is not useful advice when the strongest setting may not even exist on your plan.

Claude: first try the default Anthropic already chose

Claude introduces two controls that are easy to mix up:

model capability and effort.

Anthropic’s current Claude family spans faster models through Sonnet, Opus and Fable, while the effort menu controls how thoroughly the selected model works on a response.

The easiest starting rule is hidden in plain sight: Anthropic marks a recommended effort level as Default for each model.

For everyday work, leave it there.

Anthropic’s effort and thinking settings documentation says Low and Medium work well for routine tasks and stretch usage further. High is the default for most models and offers the best overall balance of quality and speed. Extra High is designed for long-running coding and agentic tasks and is available on Opus 4.7 and newer. Max is the most thorough option, reserved for the deepest reasoning. Higher effort takes longer and uses more tokens.

There is an important wrinkle at the top end.

In Claude, thinking cannot be turned off when using Fable 5.1 or Opus 5. On Fable 5.1 thinking is always on at every effort level, although the effort level itself can still vary. Anthropic says Fable 5.1 defaults to Medium on Claude.ai and in Claude Cowork, while Claude Code defaults it to High.

So the beginner rule for Claude is not “manually turn everything down.”

It is:

use the default until the work proves it needs more.

Fable has another wrinkle most users won’t see coming

Selecting Fable does not guarantee that every request will actually be answered by Fable.

Anthropic applies additional safeguards to Fable 5.1. In most Claude applications, cybersecurity requests caught by those safeguards can be routed to Opus 4.8, while flagged biology requests can be routed to Opus 5. Anthropic says users won’t be charged Fable prices for rerouted requests.

That isn’t a reason to avoid Fable. It is a reminder that the model name shown in the picker is not always the whole routing story. Choosely’s comparison of Fable 5.1 and Opus 5 covers where each one earns its place.

Gemini: Google’s ladder is refreshingly straightforward

Google’s current Gemini lineup gives users three fairly understandable model levels.

Flash-Lite prioritizes efficiency and speed.

Flash balances speed with stronger reasoning.

Pro is aimed at complex mathematics, coding and deeper multimodal work.

Gemini then adds reasoning levels on top.

Standard thinking is Google’s default and recommended option for most questions. Extended thinking gives complex problems more reasoning time. Deep Think is the maximum reasoning setting, but it is restricted to Google AI Ultra and requires the Pro model.

There is also a useful safety valve for people worried about exhausting their allowance.

Google says subscribers with a Google AI plan who hit a higher usage limit can continue their conversation using Flash-Lite rather than simply being locked out.

That is a good example of the underlying principle this entire guide is trying to teach: not every turn in a conversation needs the same amount of compute.

The cheapest model is not always the cheapest way to finish the job

There is a trap on the other side of this advice.

Always choosing the lightweight model can waste just as much time as always choosing the premium one.

If a cheaper model misunderstands a difficult problem three times, generates a broken approach, gets corrected and eventually forces you to move up anyway, the saving was imaginary.

For hard work, stronger models can be economical because correctness and fewer retries have value too.

The goal isn’t minimum intelligence.

It is enough intelligence.

Coding is where this starts costing real usage

The same decision becomes much more obvious inside Codex and Claude Code because usage allowances are valuable.

OpenAI currently describes its GPT-5.6 coding models as a capability and usage ladder.

For Plus users, its current published estimates for local messages per five-hour period are approximately:

These are estimates rather than guaranteed message limits. OpenAI’s Codex pricing documentation says model choice, context, reasoning, tool use, retrieval and caching can all affect usage, and the Codex pricing page adds that weekly limits may also apply.

OpenAI itself recommends switching to Terra or Luna for routine tasks when users want to extend local-message usage.

That does not mean Luna should solve every programming problem.

It means using a frontier model to perform mechanical work can consume substantially more of a limited allowance than necessary.

Claude Code makes the same idea explicit through its model choices. Anthropic’s model configuration documentation positions Haiku for simple, cost-sensitive tasks, Sonnet for daily coding, Opus 5 for complex reasoning and architecture decisions, and Fable for the largest tasks and longest autonomous sessions.

The default model also varies by account type. Anthropic currently documents Sonnet 5 as the default on Pro and Team Standard, and Opus 5 as the default on Max, Team Premium, Enterprise and the Anthropic API. Fable is not the default on any plan and has to be chosen deliberately.

For people paying attention to limited usage, coding model selection deserves a deeper guide of its own. Choosely’s breakdown of Claude Max usage limits covers how those allowances are structured.

Don’t choose by task name. Choose by difficulty

“Writing” can mean fixing a sentence or reconciling a 70-page strategy document.

“Coding” can mean renaming a button or tracing an intermittent race condition across an unfamiliar codebase.

“Research” can mean finding three straightforward options or synthesizing contradictory technical papers.

The label attached to the task tells you surprisingly little.

Instead, ask what makes the task difficult.

Does it have several interacting constraints? Is the evidence contradictory? Does the model need to maintain state over many steps? Would a wrong answer be expensive? Has the default model already failed?

Those are better reasons to increase capability than seeing a shiny model name in the menu.

The Choosely verdict

For most everyday AI use, leave the model near its normal/default setting until the work gives you a reason to change it.

Use faster models for straightforward work. Increase reasoning when the problem becomes genuinely analytical or multi-step. Reach for premium models when complexity, reliability or the cost of failure makes the extra capability worthwhile.

The model names will keep changing.

The escalation rule should survive them.

FAQ

Should I always use the newest AI model?

No. Newer and more capable models are useful when their additional capability matters. Straightforward writing, summarization and Q&A often do not require them.

Does more reasoning always produce a better answer?

No. It can materially improve difficult problems, but it also increases response time and often consumes more usage. Both Anthropic and Google explicitly recommend normal or default settings for ordinary work.

Which ChatGPT reasoning level should a Plus user start with?

Instant for straightforward work. Medium when more analysis helps. High for harder reasoning. Extra High and Pro are not currently included with Plus.

Which Claude setting should most people use?

Start with the model’s recommended Default effort setting. Raise effort when complexity justifies it. Anthropic says the lower and default levels work well for everyday tasks.

Which Gemini setting should most people use?

Google calls Standard thinking the default and says it is best for most questions. Extended is intended for complex problem solving, while Deep Think is an AI Ultra feature that requires the Pro model.

Keep your AI stack current

Choosing an AI tool is only the first decision. Model names change, reasoning levels move between plans and usage allowances get reshaped without much warning.

Stack Intelligence watches the tools you already use and surfaces the changes that actually matter to your stack.

Save your tools to Choosely and we’ll keep an eye on what changes, so you do not have to keep checking every pricing page, changelog and release note yourself.

The model picker you learned this month will not be the one you use next year.

Build my AI stack