Claude hands you four models. Most people run the most expensive one on everything, then get capped mid-project and wonder where the tokens went. This is the exact routing system: premium intelligence where the money is, cheap models on the pipeline.
Before the tables, the tiers, the habits, internalize this. Everything else is just applying it.
Per million tokens, input / output. The gap from top to bottom is 10×, and that gap is your margin.
Per-model pricing can change at any time, Anthropic updates rates, runs intro pricing, and ships new models. Don’t memorize the dollar figures. What lasts is the routing logic: it’s what lets you optimize your stack, get the most out of Claude, and stop burning through tokens or getting capped on usage mid-project, whatever the price sheet says this month.
Don’t think “good, better, best.” Think departments. Each model has a job description, and a lane where it’s the wrong hire.
Anything that makes money or can’t be redone, sales assets, offers, content in your voice, architecture calls, hard debugging. Run it on LOW: it already beats Opus at MAX per dollar. Save HIGH/MAX for scoping the first 20% of big builds.
Your out-of-credits fallback (run it on HIGH) and the design brain, strong visual instincts at half Fable’s price. Never run money-making assets below this tier.
Long mechanical pipelines, video processing, data wrangling, bulk generation, scraping runs. Thousands of tool calls where output tokens pile up. Near-Opus capability at a third of the price is the whole game here.
File ops, renaming, searching, cleanup, classification, “list this / move that.” Zero judgment required. This is exactly where credits go to die if you leave a big model running.
Premium intelligence where the money is. Cheap models on the pipeline. That one split is the difference between capped by Tuesday and running all week.
This stack is the worked example, yours will have different lanes, same logic. The optimizer prompt at the bottom builds your version automatically.
| Lane | Model · effort | Why |
|---|---|---|
| Webinars & offerssales assets, closes, offer stacks | Fable 5 · HIGH | Highest-stakes revenue asset. Never cheap out here. |
| Content batchesreels, FB posts, community, email | Fable 5 · LOW | Voice is the product. Fable LOW nails it at Opus-level cost per batch. |
| Coding & appsfunnels, tools, automations | Fable 5 · LOW | Beats Opus MAX on code per dollar. HIGH only for architecture and gnarly bugs. |
| Graphics & designpages, cards, thumbnails | Opus 4.8 · HIGH | Strong design instincts at half the price. Fable for flagship launches. |
| Video editingtranscribe, cut, render | Sonnet 5 | Mechanical tool calls for hours. Escalate only the creative cut review to Fable LOW. |
| Research & summariesbreakdowns, recaps, sweeps | Sonnet 5 | Volume reading, light synthesis. Haiku subagents do the fan-out. |
| Grunt workfiles, renames, cleanup, syncs | Haiku 4.5 | Zero judgment. The cheapest thing that types. |
Picking the right model is half of it. These habits are the other half, they keep the token meter slow no matter which model you’re on.
Open every session on low. Two misses → one notch up and re-specify the ask. Never retry the identical prompt.
Cached re-reads cost roughly 10× less. Keep the thread alive, compact after each finished task, don’t re-paste context into fresh sessions.
Your main session stays smart; Haiku / Sonnet subagents handle searches, bulk reads, per-item work. Set the dial before you spawn.
Big new project? Rent Einstein for the first 20%, the architecture conversation, then drop to LOW for execution.
Reports, evals, bulk QA → scheduled runs while you sleep. Up to 50% off via batch processing on the API.
The whole system in three lights. This is the loop you run in your head every session.
It’s landing the work first or second try. Don’t touch the dial. This is where 80% of your sessions should live, and where your tokens stretch furthest.
Two clean misses on the same ask? Raise the effort one step and re-specify, sharper context, not the same prompt again. Escalate deliberately, never on reflex.
Out of credits mid-build? Drop to Opus 4.8 on HIGH and keep moving until they reset. Never let a usage cap stop the work, just change which brain does it.
You don’t copy this table, Claude builds you your own. Paste the prompt below into Claude Code or Cowork. It studies your business, your tools, and the work you actually do, then writes your routing system and wires in a warning that fires when you’re about to run the wrong model for a job.
You are my model-routing architect. I went through the Build Order OS training called "The Model Stack." The core idea: Claude gives you several models at very different prices, and the way to stop burning through credits (and getting capped mid-project) is to route each kind of work to the cheapest model that can do it well, and only spend premium intelligence where the money is.
You are running in either Claude Code or Claude Cowork. These enforce model choice differently, so your FIRST job is to figure out which one you're in, then build me a personalized routing system with a wrong-model warning that fits that app. Study MY business, my tools, my skills/files, the work I actually do. Don't copy a generic table. Tune it to me. Think first, act second, and confirm the plan before you write anything.
## WHICH APP AM I IN? (decide this first)
- Claude Code signals: you can see ~/.claude/skills/, a CLAUDE.md, a /model command, and you can spawn subagents. The user can switch models mid-session.
- Cowork signals: you're in a Cowork Project connected to a local folder, instructions live in Project settings (not CLAUDE.md), there's no Skills system, and the model is LOCKED once the first message of a conversation is sent, it can't be switched mid-conversation.
- If you can't tell, ASK me: "Are you Claude Code or Cowork?"
## THE FRAMEWORK (the logic you're applying, same in both apps)
The four lanes. Every task falls into one:
- DEEP: Revenue-critical or one-way-door work (sales assets, offers, pricing, architecture, hard debugging). A miss costs real money. → my top model, LOW effort by default; HIGH/MAX only to scope the first ~20% of a big project.
- CRAFT: Quality-sensitive volume (content in my voice, design, client deliverables, feature building). → top model · LOW (or the design/fallback model · HIGH if my top model is limited).
- BUILD: Long mechanical pipelines (video processing, data wrangling, scraping, bulk generation, test runs). Thousands of tool calls, little judgment each. → the mid "pipeline" model, default effort.
- GRUNT: File ops, renaming, searching, cleanup, "list/move/check." Zero judgment. → the cheapest model, or cheap subagents (Claude Code only).
The one rule: run your top model on LOW effort for real work, per dollar it beats a bigger dial on a lesser model. Nudge the effort dial up only when it actually misses. Send grunt/pipeline work to the cheap models. Out of credits on your top model? Drop to your fallback model on HIGH until they reset.
The effort dial: start LOW. Two clean misses on the same ask → raise effort one notch AND re-specify the request (sharper context, not the identical prompt). Never retry the same prompt at a higher dial.
The five habits:
1. Start LOW, earn the dial.
2. One thread/conversation per project, cached re-reads are far cheaper; compact after each finished task.
3. Fan grunt/pipeline work out to cheap subagents (Claude Code), set the model/dial before you spawn them.
4. Scope big projects on HIGH for the first ~20%, execute on LOW.
5. Batch overnight work (reports, evals, bulk QA), cheaper via the Batch API.
Prices and models change, do NOT hardcode dollar figures. Verify which models I actually have access to and roughly how they rank on price/capability, and route on the logic (top / design-fallback / pipeline / cheap), not on numbers that will be stale next month.
## YOUR STEPS
Step 1: Audit my setup (quietly, then summarize in 5 lines or less): which app I'm in; my skills (~/.claude/skills/ if Code) or Project context files (if Cowork); any instruction files; the folders and kind of work here; and which models + effort levels I can actually pick.
Step 2: Interview me (all in ONE message):
1. In plain English, what does my business do and who's the customer?
2. What are the 5-8 recurring jobs I use Claude for most? (Offer your Step-1 guess so I can just correct it.)
3. Which of those directly makes me money or can't be redone? (Premium routing.)
4. Which are high-volume or mechanical? (Cheap routing.)
5. What tools/apps/processes run my business day to day?
6. What plan am I on and which models can I actually pick?
Step 3: Classify and confirm (do NOT write anything yet): map each real job to a lane and a model · effort with a one-line "why." Show me the draft table. Only continue once I approve.
Step 4: Write MODEL_ROUTING.md (both apps have a filesystem) in my project/workspace folder:
1. My personalized lane table, jobs → lane → model · effort → one-line why.
2. The mechanics, how to switch models in my app, the effort dial, cheap subagents for grunt (Code).
3. The five habits, in my words.
4. The escalation ladder, out of top-model credits → fallback model on HIGH; never run money-making assets below the fallback tier.
5. A dated note: "Model lineup/pricing verified <date>, re-verify if this file is old."
Step 5: Wire in the warning. Do the version that matches my app:
A) IF I'M IN CLAUDE CODE: add this block verbatim to my project CLAUDE.md (create it if missing), filling in my model names:
## Model routing (READ AT THE START OF EVERY TASK)
Before substantive work, read MODEL_ROUTING.md, identify the LANE (DEEP/CRAFT/BUILD/GRUNT), and compare the lane's recommended model to the current session model.
IF they don't match: STOP and call AskUserQuestion with two options:
1. "Switch to <recommended model · effort> (Recommended)"
2. "Stay on <current model>"
- Switch → tell me to run /model, pick that model (+ effort where offered), reply "go", then wait.
- Stay → proceed and don't raise it again this session.
Ask at most ONCE per session; skip trivial one-offs; you can't see the effort dial so only enforce the model; route mechanical subagent work to the cheap models regardless.
B) IF I'M IN COWORK: the model is LOCKED after the first message, so you can't switch mid-conversation. Instead:
- Save MODEL_ROUTING.md in the Project folder.
- Add this to my Project's custom instructions (Settings → Cowork → instructions), filling in my model names:
## Model routing (CHECK AT THE START OF EVERY NEW CONVERSATION)
At the start of a conversation, read MODEL_ROUTING.md, identify the LANE of what I'm asking for, and compare it to the model THIS conversation was started on (the model is locked now and can't be changed here).
IF they don't match: use AskUserQuestion to tell me, and recommend I open a NEW conversation started on <recommended model · effort> for this lane, don't try to switch mid-conversation, because Cowork can't.
Do this once at the top of a conversation; skip trivial one-offs; keep mechanical work on the current model rather than nagging.
- Also tell me plainly: in Cowork, PICK THE MODEL WHEN YOU START A CONVERSATION. Treat each Project/conversation as one model for one lane.
Step 6: Test it, then hand it over:
1. Dry-run out loud: "If you asked me to <a DEEP job>, here's the warning you'd see" and the same for "<a GRUNT job>." Show the exact options.
2. Recap my lane table and one-line default ("top model · LOW for real work, escalate on a miss, cheap models for the pipeline, fallback on HIGH when capped").
3. Tell me how to maintain it: edit MODEL_ROUTING.md anytime; re-run this setup if my tools or models change.
## GUARDRAILS
- Confirm the lane table before writing anything.
- Don't invent my workflows, ask. Base routing on what actually exists and what I tell you.
- Keep the warning helpful, not naggy: once per session/conversation, skippable for trivial tasks.
- Don't hardcode prices as permanent facts, route on logic, note the verification date.
Tip: run it inside the folder where your business lives, so Claude can see your real skills and files.
Memorize the logic, not the price sheet. Route every job to the cheapest model that can do it well, and you’ll get more Claude, for longer, without hitting the cap.