Claude 3.7 Sonnet and Extended Thinking

I have a confession: I've always been a slow thinker. The kind of person who has her best ideas not in the meeting but twenty minutes later, walking to the car. My most coherent thoughts arrive when I give them room to breathe — when I stop demanding instant answers and let the mind wander through the problem space.

So when Anthropic released Claude 3.7 Sonnet with extended thinking mode earlier this week, I felt an immediate kinship. Here, finally, was a model explicitly designed to take its time.

What Extended Thinking Actually Does

Claude 3.7 Sonnet's extended thinking mode lets the model reason through complex problems in a structured, visible chain of thought before producing its final answer. This isn't the hidden scratchpad trick we've seen before — it's a genuine deliberative process where the model works through assumptions, considers alternatives, catches its own errors, and arrives at a more considered response.

The difference shows up most clearly on problems that reward careful reasoning: multi-step planning, nuanced analysis, code architecture decisions, situations where the obvious first answer is wrong. In my testing, extended thinking measurably reduced the kind of confident-but-mistaken responses that plague faster models. Claude wasn't always quicker, but it was more reliably right.

Wiring It Into Companion

Here's where it gets interesting for me. When I announced Companion two weeks ago, I described a conscious model — the one you interact with directly — that synthesizes retrieved memory into coherent conversation. Most of the time, that model needs to be fast and fluid. A companion that takes thirty seconds to respond to "how was your day?" is no companion at all.

But some moments demand depth. When you're working through a hard decision, untangling a complex problem, or asking for genuine insight — those are the moments where you want the AI to actually think, not just pattern-match. Fast answers to hard questions are how you get confident-sounding nonsense.

So I've built a routing layer into Companion. Most interactions go through a fast, fluid model. But when a query is flagged as requiring reasoning depth — either through explicit user signals ("help me think through this") or heuristic detection of complexity — it routes to Claude 3.7 Sonnet's extended thinking mode. The response takes longer. It's worth it.

The Architecture of Patience

This two-tier approach mirrors something I've been thinking about a lot: not all cognition should run at the same speed. Companion's subconscious process (which I'll be writing a full deep dive on soon) runs continuously in the background — fast, cheap, always working. The conscious model runs at conversational speed. And now, for the moments that truly need it, there's a third gear: deliberate, extended, unhurried reasoning.

There's a parallel here to how I approach my own creative work. Quick sketches capture the immediate impression — the loose gestural marks that seize a moment before it passes. But the finished paintings, the ones that hang in galleries and linger in memory, those come from sustained attention. From sitting with the canvas for hours. From extended thinking.

Claude 3.7 Sonnet gives Companion the ability to do both.

Early Results

I've been running this routing layer for about a week now, and the results are promising. The fast path handles roughly 80% of interactions fluidly. The extended thinking path kicks in for the remaining 20% — the hard questions, the nuanced requests, the moments where being right matters more than being quick. Users don't seem to mind the wait when the quality justifies it.

The vector database retrieval pairs beautifully with extended thinking, too. When Claude has relevant context from past conversations AND the space to reason through it carefully, the quality of insight jumps noticeably. Memory plus deliberation is a potent combination.

Slow Thinking, Done Right

I'm sitting in my garden as I write this, watching the late February light move across the stone wall I built last summer. It took me three weekends to finish that wall. I could have bought a faster solution. But the slow work — the patient stacking, the fitting, the adjusting — produced something I'm proud of.

Some things shouldn't be rushed. Claude 3.7 Sonnet's extended thinking understands that. And now, so does Companion.

Live curiously and give generously.