A 3am theory about cognitive architecture, Kahneman, and why I always reach for Sonnet.
Kahneman's Two Systems
Daniel Kahneman spent decades studying how humans make decisions. His conclusion, laid out in Thinking, Fast and Slow, is that we don't have one brain — we have two systems running in parallel:
System 1 is fast, automatic, instinctive. It pattern-matches. It reacts. It's the part of you that catches a ball before you consciously decide to reach for it. It doesn't deliberate — it fires.
System 2 is slow, deliberate, effortful. It's the part of you that works through a logic puzzle, reads a contract, or plans a difficult conversation. It's expensive — metabolically and cognitively. You can't run it all day.
Most of cognition, Kahneman argues, is System 1. System 2 is reserved for when it counts.
The Third Layer Kahneman Didn't Name
Neuroscience has a different framing — one that predates Kahneman and goes deeper.
Paul MacLean's triune brain model describes three evolutionary layers:
The reptilian brain — brainstem and cerebellum. Pure survival. Reflexes, breathing, heart rate. No reasoning. Just reaction.
The mammalian brain — the limbic system. Emotion, memory, social bonding, creativity. The part of you that feels, connects, and imagines.
The human brain — the neocortex, especially the prefrontal cortex. Abstract reasoning, planning, language, deep thought. The thing that makes us distinctly us.
The interesting part: these layers don't replace each other. They stack. They run in parallel. The neocortex doesn't turn off the limbic system — it converses with it.
Three Models. Three Layers.
Here's the theory:
Haiku is the reptilian brain.
Fast. Cheap. Reflexive. It pattern-matches and responds. It's System 1 made silicon. You use Haiku for the things that need to happen instantly — classification, tagging, quick lookups, high-volume inference at the edge. It doesn't deliberate. It fires.
Sonnet is the mammalian brain.
And this is where it gets interesting.
Sonnet is not the compromise between fast and slow. It's something qualitatively different. The limbic system isn't a weaker neocortex — it's where creativity lives. Emotion. Connection. Flow.
Sonnet is where you go when you need to think with feeling. When the problem isn't purely logical and it isn't purely reflexive. When you're writing, exploring, building, ideating. When you're at 3am and something is trying to surface.
Opus is the neocortex.
Slow. Expensive. Deliberate. It reasons through problems that require holding multiple competing ideas in tension. You bring Opus when the stakes are high and the problem is genuinely hard — architecture decisions, legal analysis, long-horizon planning. It's System 2. You can't run it on everything; it costs too much.
Why This Matters for How You Work
If this model is right, the choice of which model to use isn't just a cost-performance tradeoff. It's a cognitive mode selection.
Routing a creative problem to Opus is like asking a team of senior engineers to brainstorm your product name. You'll get a technically rigorous answer when you needed a generative one. You'll get analysis when you needed flow.
Routing a complex architecture decision to Haiku is the inverse mistake — you're asking your reflexes to do your planning.
The instinct to "always use the smartest model" misses this. Bigger isn't always better. It's about matching the cognitive mode to the task.
| Model | Brain Layer | System | Use Case |
|---|---|---|---|
| Haiku | Brainstem / Cerebellum | System 1 | Extraction, tagging, classification, high-volume inference |
| Sonnet | Limbic System | Flow state | Writing, coding, building, exploring, 3am crystallization |
| Opus | Prefrontal Cortex | System 2 | Architecture, legal, long-horizon planning, high-stakes reasoning |
A Note on the Naming
I don't know if Anthropic intended this. The official explanation is capability tiers: Haiku is fastest and cheapest, Sonnet is balanced, Opus is most capable.
But the poetic forms map. The cognitive layers map. The systems map.
Maybe that's coincidence. Maybe it's the kind of thing that happens when a company thinks carefully about what they're building.
The Practical Takeaway
Next time you open a new conversation, don't default to Opus because it "feels safer." Ask what mode you're in:
- Classifying, reacting, extracting at scale → Haiku
- Creating, exploring, building, flowing → Sonnet
- Reasoning through hard, high-stakes, complex problems → Opus
And if you find yourself reaching for Sonnet at 3am when something is trying to crystallize —
You already know why.
Written at 3:30am. First draft always beats the edited version.
— Rahul
I'm building knowledge-worker — a local-first knowledge graph that gives AI assistants durable, provenance-backed memory across sessions. If you've ever re-explained the same context to Claude three times in a week, it might be for you.