Orbit

Low, High, Max, Ultra... how hard should my AI think?

Chris Rupp

You ask your agent to fix a bug and it answers in like two seconds… and it’s wrong. Then you notice there’s a little setting sitting right next to the model that says effort. Low, medium, high, max. What does that even do, and why would anybody pick low?

Effort is how long it thinks before it talks

That’s really all it is. On low the model answers right away, first thing that comes to mind. On high it thinks first… it makes a plan, checks its own work, catches its own mistakes, and then it answers.

The catch is that thinking isn’t free. It takes longer, and you pay for it, cuz thinking is made of tokens too, even the ones you never actually see on your screen.

Wait, isn’t that what a reasoning model is?

You’ve probably heard people say “use a reasoning model” like it’s some special kind of AI. And it used to be, they were separate models with their own names. Now Claude and GPT can both do it, and effort is just the dial in between quick answer and deep thinking.

Match the effort to the job

And here’s the part I wish somebody told me sooner… on easy jobs more effort doesn’t get you a better answer, it just gets you the same answer slower and more expensive.

Effort is also a separate dial from model size. You can turn it up on a small model, or down on a big one. To change it in Claude Code type /effort, and in Codex it lives in /model.

Then there’s Ultra

For the really big jobs both of them have a team mode now. In Codex you pick Ultra under More reasoning, and in Claude Code you turn on ultracode in /effort, or just type ultracode in your prompt. Instead of one agent thinking longer, the job gets split across helper agents all working at the same time.

That burns through your limits a lot faster though, so save it for big jobs that actually split into parts, not your button color change lol.

How I stopped babysitting the dial

I kept forgetting to crank it up for the hard stuff and back down for the easy stuff, so in Orbit, the program I built to run my agents, each model tier gets its own effort level. Low, Medium, High and Ludicrous each get a model and an effort once, every prompt in the queue gets a tier, and quick jobs stay quick while the hard ones get time to think. If one prompt needs something different you can override the effort on just that one. Orbit is in Alpha and free right now if you wanna try it: https://orbitdev.app

All of these words, and the ones that come next, are in the Learning Atlas if you want to go deeper.

So be honest… do you leave effort on max all the time, or do you actually change it per job?