More options means slower decisions. Constraint creates execution. Today we unpack this principle through free AI models.
Let me ask you something. You walk into a restaurant and the menu has 200 items. How long will it take you to order? Now imagine the menu has just three items.
The answer is obvious. More menu = longer decision. We know this in our bodies. Strange, then, that in front of AI tools the knowledge vanishes. "28 free AI models!" and suddenly we feel we must try them all — and a month later, none have been used.
This essay shows you how to escape the trap. Even if you've never used an AI tool, you can follow along. We'll go slowly. Today's example is AI models, but the principle is identical for exercise, books, apps, anything. Too many means you can't pick. Constraint creates execution. Three years from now, when the AI landscape is completely different, the spine of this essay still works.
More options feels good, but three bad things happen as they multiply.
First, deliberation time balloons. Choosing from 3 vs from 30 — which is longer? Obviously 30. But not just 10× longer. Much longer. Comparing 3 takes 3 comparisons. Comparing 30 takes 435. A serious weekend gone, considering nothing is built.
Second, satisfaction drops after the pick. From 30, you wonder "maybe that other one was better?" From 3, that thought never appears. More seen means more abandoned.
Third, and most important — you don't start. Chasing perfect means endless delay. Nothing happens during delay. But learning comes from doing, not from comparing. This is the whole essay.
I was slow to learn this. Every new tool: review videos, Twitter, blog comparisons, and the weekend was gone. Monday morning — "what did I actually use?" — nothing. All weekend comparing, zero using. Grab three at random and spend one day actually using them — you'll learn more than from a full week of comparing twenty-eight.
You don't do this in familiar territory. Three movie options? Pick one. 28 printers? Don't compare all. But sit down to AI and something changes.
It's unfamiliarity. In new territory you defensively try to compare everything. "Don't want to pick wrong." But the comparison itself is the biggest mistake. Time lost, decision stuck, starting blocked. Doing nothing is the most expensive choice.
So today's common sense to recover is one line — "Three is enough." In any new field, start with three.
Concrete now. OpenRouter is a service that aggregates AI models from many companies in one place. As of April 2026, it hosts 28 free models — Gemini, Llama, DeepSeek, Qwen, Mistral.
The number alone is exciting — "28 free!" But sit down to choose and days disappear. Spec comparisons, benchmarks, reviews. Evening comes, "I don't know yet," laptop closed. Nothing used.
The way out was simple. I defined what I wanted to do first, not the tools. Not specs. Uses.
Once three uses were defined, model selection was automatic. One per use. Three total. Done.
| Use | Model | Speed | Notes |
|---|---|---|---|
| Fast answers | Gemini 2.0 Flash (free) | 1-2s | Follows instructions cleanly |
| Code | DeepSeek Chat (free) | 3-5s | Stable code quality |
| Long documents | Llama 3.3 70B (free) | 5-10s | Good context handling |
Gemini Flash is my morning-open. "Organize today's schedule," "draft a reply to this email" — anything that needs 5-second turnaround. 50+ times a day.
DeepSeek only opens for code. A Python function, a React refactor. Noticeably more stable than Flash on code.
Llama 70B only for PDFs or long docs. "Extract the key points from this 30-pager" — Llama wins. Slow, but depth with speed-tradeoff. 5 seconds beats a 10-minute manual summary.
The three personalities are this clear for one reason — I lived with only them for a month. People who touched all 28 lightly never know the real personality of any. I can pick blindfolded. Depth comes from narrowing.
These three handle 90% of my free-AI work. Reviewing 25 more for the remaining 10% is waste. Here comes the second aha:
Three isn't everything. Three is the start.
Handle 90% first. After the limits become obvious, add a fourth. One at a time. Trying to absorb 28 at once means mastering none.
# Sign up at OpenRouter, get API key, add to ~/.zshrc:
export OPENROUTER_API_KEY="sk-or-v1-..."
export ANTHROPIC_BASE_URL="https://openrouter.ai/api/v1"
export ANTHROPIC_API_KEY="$OPENROUTER_API_KEY"
# Fast answers
claude --model "google/gemini-2.0-flash-exp:free"
# Code
claude --model "deepseek/deepseek-chat:free"
# Long documents
claude --model "meta-llama/llama-3.3-70b-instruct:free"
Three minutes to set up. Your monthly AI cost approaches zero.
The critical part is staying on these three for a month. One week isn't enough. Don't rotate every three days. Resist the "maybe that other one…" voice for three weeks and something strange happens in week four — the tool's personality starts fitting your tasks. Moments of fit accumulate, and they become feel. Feel is expertise.
A month in, two things change. First, you stop hesitating in front of the tool — "this is the A model" your body says before your head does. Second, more important — "start with three" becomes a habit across domains. You stop adding 28 books to the cart — you buy 3 and read for a month. You stop trying every workout program — you run 3 moves for 6 weeks. Stack a year of this habit and your execution time is way ahead of the comparing crowd.
That's the entire principle.
Too many means you can't pick. Perfect blocks starting. The answer is constraint. Start with three, live with them a month, widen one at a time when limits show.
Three words to carry — Fast. Code. Long. Three uses. Three tools. Uses first, tools follow.
It isn't only for AI. Starting exercise? Pick three movements. Programming? Three languages. Apps? Three per function. Constraint creates execution. Three years from now, we won't know what Gemini and Llama are called. The principle keeps working. When some new company ships 50 models, you'll still pick three. Because it's not the tool's name that makes the selection — it's the definition of use. Technology changes. The principle doesn't.