​

Most of us are using AI the way we used to use Google.

Open it. Ask it something. Get an answer. Close it.

The problem: frontier models aren't Google. They're specialists on a meter. Using GPT-5 or Claude Opus to draft a routine email is like hiring a senior architect to hang a shelf. The work gets done. The cost doesn't match the job.

And on the other end, using a lightweight model to build a complex agent or write a nuanced client strategy is like handing a hammer to someone who needed a blueprint first. Also wrong.

The fix isn't using AI less. It's using the right model for the right kind of work.

Introducing: The Tier Stack

The Tier Stack is a three-level mental model for matching work to model. Here's how it breaks down:

Tier 1: Frontier (Claude Opus, GPT-5, Gemini Ultra)​
Use this when you're building something new. Designing a workflow. Writing a prompt system from scratch. Thinking through strategy. Creating a template or SOP that will get reused. This is the architect tier. You're paying for judgment, nuance, and the ability to handle ambiguity.

Tier 2: Mid-tier (Claude Sonnet, GPT-4o, Gemini Flash)​
Use this for execution. Running the workflow you built. Drafting inside an established framework. Summarizing, formatting, iterating. This is the builder tier. Fast, capable, and a fraction of the cost.

Tier 3: Lightweight / Specialized​
Use this for repetitive, structured, or templated tasks where the output format is already defined. Data extraction. Classification. Formatting checks. This is the assembly tier. When the scaffold is already up, you don't need the crane anymore.

The workflow that actually makes this work:

1. Use Tier 1 to build the scaffold: the prompt, the system, the template
2. Validate the output at Tier 1 until you trust it
3. Hand execution to Tier 2 or Tier 3

That's it. One decision loop. The frontier model earns its cost by building something that runs without it.

What the Caveman plugin taught me about this

Someone shared a tool this week called Caveman (a Claude Code plugin that strips output down to the minimum tokens needed for technical accuracy). A normal response might run 69 tokens. The Caveman version: 19. Same information. No loss in technical precision.

That's not just a neat trick. It's proof of a principle: verbosity is a habit, not a feature. And when you stop paying for it at scale, the savings compound fast.

The same logic applies across your stack. You're not paying for intelligence when you use a frontier model on routine work. You're paying for overhead.

Here's what people miss about their AI subscriptions:

Most companies now have access to more AI capability than they're using well. Microsoft 365 Copilot is running quietly inside tools most teams already have open. Codex is running skills and automations that most operators haven't built yet. The gap isn't access. It's strategy.

The professionals who understand cost and model economics right now (who can say "this task is Tier 1 work, this one is Tier 3") are building a skill that's genuinely rare. The AI job market data backs this up. Cost and token economics shows up on senior-level job postings more than almost any other AI-specific skill.

You don't need a new subscription. You need a framework for the ones you already have.

What are you currently using your frontier model credits on?

Reply and let me know!

I'd love to see what cool projects are out there I'm missing and if this Tier Stack is helpful.

​