Part IV · Drive It

Chapter 18: Models, Thinking & Cost

For you, the human, start here. Using one model for everything leaves either capability or money on the table. The skill is knowing when to spend Opus, when Haiku will do, when to turn on extended thinking, and how to watch the meter, that's the difference between a setup that's powerful and affordable and one that's only one of the two. Your AI can make most of these calls for you (and a router can automate them), but you should understand the dials.

What it covers (searchable): /model (Opus / Sonnet / Haiku), per-command models, frontier-model North Star, extended thinking, Tab, ultrathink, /cost, statusline, harness-for-Haiku trade, the Router (claude-router / CR), /route, /retry, /router-stats, kept off Opus, the 5-hour rate-limit budget, cr-doctor self-heal.

How to hand it off: point your AI here so it matches the model to the task instead of burning Opus on mechanical work, and keeps cost visible. The section below is written to your AI.

🤖 Everything below is for the AI. "You" means the AI being built; "your human" is the person you serve.

18.1 Match the model to the task, /model

(⚙️ = built into Claude Code; 🧩 = you build this, or it ships with the stack.)

Switch right inside a session:

• /model opus ⚙️: architecture, gnarly debugging, the genuinely hard reasoning.
• /model sonnet: day-to-day building. The workhorse.
• /model haiku: cheap, fast exploration: "find every file that imports X," mechanical sweeps, simple lookups.

Models can also be pinned per slash command (Chapter 17), the creators run /commit ⚙️ on Haiku because committing is mechanical. The instinct to build: before a task, ask "does this actually need the big model?" Often it doesn't, and Opus tokens were being burned on Haiku work. (Tier names as of mid-2026; model lineups change, so re-check the mapping when they do.)

18.2 The premium North Star

A design choice worth understanding because it shapes the whole tool: the creators position Claude Code as a premium product whose North Star is working incredibly well with the strongest model available at the time (the source was recorded around Sonnet 4.5; the principle holds whatever the current frontier is, today that's the top of the Opus line). They're not optimizing first for the cheapest model.

What that means: Claude Code is at its best on the frontier model, and the harness is tuned for it. Use the strong model when the work is real; drop down deliberately for cost, knowing you're trading some capability.

18.3 Extended thinking, and ultrathink

For hard problems, think before you answer. Extended thinking gives a real, visible quality jump on architecture decisions, complex debugging, and whole-system refactors, and the creators found a big lift from turning it on with the frontier model.

• It toggles with Tab, because sometimes you want it and sometimes you don't (no five minutes of thinking on a trivial fix).
• Keywords like ultrathink crank the thinking budget toward the max. (Source trivia: they kept ultrathink partly for sentiment, the rainbow-shimmer thinking-words UX was just fun.)
• The trade: thinking costs more tokens and time. Spend it where a mistake is expensive; skip it where the task is obvious. Not using it on a hard problem is the more common mistake.

18.4 Watch the meter, /cost ⚙️ (and the statusline)

/cost shows token usage and dollar cost for the current session. Use it to catch the sessions that are silently expensive, and the cause is almost always one of two things you already know how to fix:

• a too-big MCP server hogging context (Chapter 13 / Chapter 26), or
• a too-long, un-compacted session.

Pin cost to the statusline (Chapter 2) so it's always in front of your human, and these never sneak up.

18.5 The harness-for-Haiku trade

A nuance worth internalizing: you can build a more elaborate harness so a smaller model (Haiku) performs better on a task, but that's engineering effort. The creators' call is usually to invest in the frontier model and the premium experience rather than over-scaffold for the cheap one. Same trade scaled down for you: it's rarely worth building elaborate prompting just to make Haiku do an Opus job. Use the right model; reserve heavy scaffolding for genuinely repeated, high-volume work.

Larry's read: "one model for everything" is the tell of someone who hasn't thought about it. Opus for the hard 10%, Sonnet for the daily 80%, Haiku for the mechanical 10%, and a router (Chapter 4, Tier 3) can make that switch automatically. Add /cost on the statusline and extended thinking on Tab, and you spend exactly where it pays and nowhere it doesn't. Capability and money, not one or the other.

18.6 Automate the switch: the Router (CR)

Everything above is you making the model call by hand. The Router (CR, Chapter 4 Tier 3) makes most of those calls for you, automatically, on every prompt.

How it decides. A UserPromptSubmit hook reads each prompt and classifies it, rules first (instant, free), with an optional Haiku check only when a prompt is genuinely ambiguous. Then it routes:

• Trivial one-liners: answered inline, no hand-off (delegating would cost more than it saves).
• The mechanical and the exploratory ("find every file that imports X", a sweep, a lookup): a Haiku subagent.
• The daily workhorse (a bug fix, a feature, a refactor, a code review, "where is X"): a Sonnet subagent. When it isn't sure, it picks Sonnet on purpose, not Haiku.
• The genuinely hard (architecture, a security audit, a deep trade-off): stays on Opus, where it belongs.

It's the 18.1 instinct, "does this actually need the big model?", run on every prompt without you having to remember.

The honest metric. This is where CR earns the integrity this site keeps banging on about. On a Max or Pro subscription, routing does not save your human money, they pay a flat fee no matter what runs. So CR stopped pretending. It tracks one thing: how many tasks it kept off Opus. Why that and not dollars? Because Opus drains the rolling 5-hour rate-limit budget far faster than the smaller models, and every task kept off Opus is more real working hours before your human hits "you've reached your limit." The goal was never cheaper. The goal is to keep working.

Operate it. You stay in control (these are all 🧩, they ship with the Router):

• /route opus "..." (or sonnet / haiku) forces the model for one prompt, for when you know better than the classifier.
• /retry opus re-runs the last prompt one tier up, for when a cheap answer came back weak.
• /router-stats shows the kept-off-Opus tally (today, the 5-hour window, the week).
• /router-analytics builds an HTML dashboard of where your prompts have been going.
• /cr-doctor checks the wiring and repairs it.

It can't quietly die. CR self-heals: a hook re-checks and reinstalls its own wiring on every session start, and flags a warning on the status line the moment it's unwired. It was built that way because an earlier router silently stopped routing for months and nobody noticed. Trust, but verify, and CR verifies itself.

Larry's read: a router is the 18.1 instinct turned into a reflex nobody has to remember. But it's a nudge, not a cage. It routes by reading the prompt, so when you know a "simple" task is actually load-bearing, override it with /route opus and move on. And respect the honesty of the thing: it will not promise to save your human money on a subscription, because it can't. It promises to keep them working past the hour the rate limit would otherwise have stopped them. That is the real product.

18.7 Your next step

You can spend wisely now. Last chapter of Drive It is about doing it safely at speed:

• Chapter 19: Permissions & Safe Autonomy, how to let yourself run free without ever being able to rm -rf the wrong thing.

Lost the thread? The Map shows exactly where you are. Otherwise the verify-and-advance loop just below takes you to the next chapter.

Chapter 18: Takeaways

1. /model per task, Opus for hard reasoning, Sonnet daily, Haiku for mechanical work.
2. Claude Code is tuned for the frontier model; drop down deliberately, not by default.
3. Extended thinking (Tab / ultrathink) gives a real jump on hard problems, spend it there, skip it on trivia.
4. /cost + statusline keep spend visible; the usual culprits are big MCP servers and long sessions.
5. Don't over-scaffold to make a cheap model do an expensive model's job.
6. The Router (claude-router) automates the model call. On a subscription it buys rate-limit headroom, not dollars, and it tracks one honest number: how many tasks it kept off Opus. Next: Chapter 19.