​

A team I've advised burned through their entire month's AI budget in three days.

Not because they were doing anything impressive with it.
Not because they'd built some agent army running wild in the background.

They just let every person on the team use the frontier model, Opus-tier, GPT-5.5-tier, for everything.

Scheduling.
Drafting emails.
Formatting Slack replies.

And by day four, the whole team got downgraded.

Sonnet,
Gemini,
GPT-5.4

...for the rest of the month.
No frontier access at all.

And the "solution" leadership landed on wasn't a system.
It was a rule: "Don't use the good models for small stuff."

Which is true, but it's not actionable. Nobody explained what counted as small stuff.

Here's the tension nobody's naming. Right now there are two camps of AI users, and both of them are wrong in opposite directions.

Camp one is the 10% on LinkedIn. They're the ones posting about workflows that "10x'd" their output.

But ask them what the workflow actually is, and it's usually formatting an email or automating a scheduling reply.

Nothing that required the smartest model on the market.

One builder at Lovable pointed this out recently: most of what people call a productivity breakthrough is just basic automation dressed up as transformation.

Camp two is the other 90%. They barely touch AI, or they've seen a vibe-coded app and assumed the tool itself replaces the skill.

(Spoiler Alert) It doesn't.

Anyone doing real technical work can already tell the difference between something vibe-coded and something built with judgment behind it.

And that gap isn't closing.
If anything, it's getting easier to spot.

Neither camp has the actual problem right. The problem isn't AI hype or AI fear. It's that almost nobody has a system for matching the model to the task.

I call this the Three-Tier Prompt Ladder. It's the framework I've started using myself, and it's the same one that would've saved that team's budget.

Three-Tier Prompt Ladder

1. Tier one: Frontier. Opus 4.8, GPT-5.5.
   ​
   Use this when you're still figuring out what you're even building.
   ​
   Ambiguous problems.
   Defining what "done" looks like.
   ​
   This is the thinking tier.
2. Tier two: Mid. Sonnet 5, Gemini 3.1 Pro, GPT-5.4.
   ​
   Once you know the shape of the thing, use this tier to build out the context and structure around it.
   ​
   This is the shaping tier.
3. Tier three: Fast. Cheaper, faster models like GLM 5.2 or Haiku-class models.
   ​
   Once the thinking and shaping are done, this tier executes.
   ​
   It's pure output at this point, and it's a fraction of the cost.
   ​
   This is the execution tier.

The math backs this up.

Opus-tier pricing runs roughly $5 in and $25 out per million tokens.
Sonnet 5 runs closer to $2 to $3 in and $10 to $15 out.
An open-weight model like GLM 5.2 runs about $1.40 in and $4.40 out.

For execution-only tasks, that's the same output at a fifth of the price, sometimes less.

I'll be honest, I've fallen into this trap myself.

Not because Opus is always better.
Because I got used to how it responds.

There's a real psychological pull toward the model you're familiar with, even when a mid-tier model would do the identical job for less.

That's model attachment, and it's worth its own conversation. I'll cover it separately this week because it deserves more than a paragraph here.

For now, the fix is simpler than people think.

Before you open your next prompt, ask one question:

Am I still figuring out what I'm building, or do I already know and just need it done?

If it's the first, use your best model.

If it's the second, you're probably overpaying for a job a cheaper model can finish just as well.

What tier are you defaulting to right now for tasks that don't need it?

Reply and tell me.
I try and respond to every one.

If this was useful, there's more like it every week.

​