Guide | Settings

GPT-6 effort levels explained

One setting controls how hard GPT-6 thinks before it answers. Light, Medium, High: what each is for, when to move off the middle, and why OpenAI says the setting matters less than your prompt.

Start here

One setting decides how hard it thinks

GPT-6 Astra has an effort setting. It controls how much the model thinks before it answers you. Lighter effort is faster and cheaper. Higher effort thinks longer and does better on hard problems.

In the ChatGPT app the three choices are Light, Medium and High. The model also adjusts on its own inside whatever level you pick, using less thinking on simple questions and more on complex ones. So the level is a ceiling, not a fixed speed.

Here is what each one is actually for, and when it is worth moving off the middle.

Light, Medium, HighWhat each one doesWhen to move off MediumSeptember 2026
The levels

What each one is actually for

Three choices in the app. That covers everything most people do.

01
Light

What it's for: efficient thinking with only a small wait. Built for work involving tools, planning, search, or multi-step decisions where you want speed and low cost more than depth.

Reach for it when
  • You are drafting and you will edit anyway
  • You are pulling facts out of something you already have
  • You are sorting, tagging, or categorizing
  • You want an answer in seconds, not a better answer in a minute
02
Medium

What it's for: the default. For when quality and reliability matter and the task involves planning, real reasoning, and judgement. It balances waiting, quality and cost.

Reach for it when
  • It is most of your week
  • The output goes to a client, a customer, or your accountant
  • The job has several steps and an order that matters
  • You are not sure. Start here.
Whose call this is

Ruben Hassid keeps his on Medium after using Astra every day for a week, and calls it the best mix of speed, cost and thinking.

OpenAI does not publish a best setting for ChatGPT. Medium being the documented default is as close as it gets.

03
High

What it's for: hard reasoning, complex problems, deep planning, and high-value work where quality matters more than how long you wait.

Reach for it when
  • The answer is expensive to get wrong
  • You have tried Medium and the reasoning is where it fell down, not your instructions
  • It is a decision rather than a draft
  • You can afford to wait
The part that matters most

Effort is a tuning knob, not the fix

This is OpenAI's own guidance and it is worth more than the rest of this page: treat the effort setting as a tuning knob, not the main way to get better answers. Task clarity, constraints, and saying what the output should look like matter more than the level you pick.

So if an answer comes back wrong, turning the dial up is the second thing to try. The first is to say what done looks like, what each attached file is for, and what shape you want the answer in.

Where to start

Leave it on Medium. When something comes back wrong, fix the prompt before you touch the setting. Move to High only when you have ruled the prompt out.

Modes

The other picker, and what it is

Effort is how hard it thinks. Mode is what kind of work you are asking for. Ruben Hassid works only in Work or Codex mode with Astra.

Which modes you see depends on your plan, and OpenAI's public model documentation does not list ChatGPT plan eligibility, only API usage tiers. So open your own menu and look at what is actually in yours rather than trusting any list, including this one.

One more thing

If you go reading OpenAI's developer docs

Only relevant if you build with the API. Skip this if you use ChatGPT in the app.

The developer documentation uses different names for the same dial. What the app calls Light is low there. It also documents levels the app does not show you: xhigh for deep research and long autonomous runs, and max for the most complex tasks. There is a none level too, which Astra does not support and which returns an error.

If a name in the docs does not match what is on your screen, your screen is right.