Welcome back to my ongoing series about discovering AI tools that everyone else has been using since forever. If you’ve been following along, you know I’ve covered GitHub Copilot’s custom instructions in Part 1 and autonomous coding agents like OpenCode in Part 2.
Today, I want to talk about something that doesn’t get nearly enough attention in my opinion: model selection. Or more bluntly: why using the most expensive, powerful AI model for every single task is not just wasteful – it’s irresponsible.
The Problem
Here’s a scenario we witness in video, blogs and social media occasionally: Someone opens ChatGPT, has the fancy new reasoning model (the one that thinks really, really hard about things) preselected, and uses it to write a two-sentence email.
Or they fire up Claude Code with their awesome Opus model to fix typos in a README or use it to generate a simple “Hello World” script.
Does it work? Sure. Is it completely ridiculous? Absolutely.
It’s Not Just About Your Token Budget
When you pick an AI model, there’s an obvious cost: your subscription limits. Most AI tools give you a certain number of tokens per day with the expensive models. Burn through those, and you’re either waiting until tomorrow or paying extra.
But here’s what a lot of people don’t realize: every AI query consumes real energy. Not metaphorical energy. Actual electricity. The kind that comes from power plants and has a carbon footprint.
Running a massive reasoning model or the latest frontier model for trivial tasks doesn’t just waste your included tokens – it wastes energy. Lots of it. And when millions of people do this multiple times a day, that adds up to a significant environmental impact.
I’m not trying to guilt-trip anyone here. I’m just saying: we have a responsibility to use these tools thoughtfully. Just because you can use the biggest, smartest model for everything doesn’t mean you should.
The Right Tool for the Job
Think about it like this: you wouldn’t use a sledgehammer to crack a nut. And you probably shouldn’t use a cutting-edge reasoning model to write a simple email.
Different tasks need different levels of intelligence:
Simple tasks – Writing emails, fixing typos, generating boilerplate code, answering straightforward questions → Use a lightweight, fast model (GPT-5-mini, Claude Haiku, or similar)
Complex tasks – Architectural decisions, debugging tricky issues, designing systems, solving novel problems → Use the heavy hitters (o1, Claude Opus)
For everything in between→ Use a mid-tier model (GPT-4o, Claude Sonnet)
The pattern is simple: match the model to the task. Don’t overcomplicate it.
How This Applies to Coding (and Everything Else)
In Part 2, I talked about OpenCode’s plan mode vs. build mode. This is where thoughtful model selection becomes really powerful.
Let’s say you’re implementing a complex feature or doing a major refactoring. This is exactly the kind of work where you want a smart model. The architecture matters. The approach matters. Getting it wrong could mean hours of wasted effort, tokens and energy.
So here’s one strategy to approach this:
Use a Smart Model for Planning
When you’re in plan mode – analyzing the codebase, designing the approach, thinking through edge cases – go ahead and use a powerful model. This is where intelligence pays off. You want the AI to really understand the problem, consider different approaches, and create a solid implementation plan.
This is expensive. But it’s worth it. Because a good plan saves you from going down the wrong path.
Use a Cheap Model for Building
Once you have a clear, detailed plan, the actual implementation is often straightforward. You’re just following the blueprint. Writing the code, updating files, making the changes – this doesn’t require genius-level reasoning.
So switch to a cheaper, faster model for build mode. It’ll follow the plan just fine, but it won’t burn through your token budget or consume unnecessary energy doing it.
Closing Thoughts
I’m still learning this myself. My default instinct is often to reach for the mid-tier model, because it feels reasonable. But I’m trying to be more intentional about it.
And here’s the thing: most of the time, the cheaper models work just fine. Sometimes they’re even better – faster responses, less overthinking, more direct answers.
So my challenge to you (and to myself) is this: next time you’re about to use an AI tool, pause for half a second and ask yourself: “Do I really need the premium or mid-tier model for this task?”
If the answer is no, dial it back. Your future self – and the planet – will appreciate it.
Remember: just because you’re late to the AI party doesn’t mean you can’t help set better standards for how we use these tools going forward.