Build your AI stack

Why most people should run more than one model

The single-model habit is a subscription artefact, not a workflow decision. If you pay one provider $20 a month you will use that provider for everything, including the jobs it is worst at. On the API the calculus is different: you are billed per token, so keeping four models on hand costs nothing until you call them, and the gap between the best and worst model for a given task is far wider than the gap between two frontier models on a leaderboard.

The practical version of this is a stack: a named model for each recurring job, chosen once, so you stop re-litigating the decision every time you open an editor. Four slots cover most people.

The four slots worth filling

The daily driver

The one you reach for without thinking: interactive coding, drafting, everyday questions. It should be fast enough that you do not mind waiting and cheap enough that you do not count the calls. This is a mid-tier model, not a flagship — the ceiling matters far less here than latency and price, because you will make more calls to this slot than to the other three combined.

The ceiling model

Reserved for work where a wrong answer costs more than the tokens: a refactor across unfamiliar code, a legal or financial read, an agent loop that runs unattended. You might call it a few times a week. Pick on capability alone and ignore the price — at that volume the price is not the constraint.

The volume workhorse

For anything that runs thousands of times and gets checked before it ships: classification, extraction, tests, boilerplate, bulk rewriting. This is the only slot where cost per million tokens is the deciding number. It is also the slot people fill worst — a model too weak to do the job is not cheap, it is free output you have to throw away.

The long-document reader

Whole repositories, transcripts, contracts, research sets. Chunking is where quality quietly dies, so the question is whether the input fits in one pass. If your inputs never exceed a few thousand tokens you do not need this slot at all.

How to choose each one

Score the slot, not the model. For the daily driver, sort by speed and read the price column second. For the ceiling model, sort by the capability axis that matches your work and ignore everything else. For the workhorse, start from the cheapest model that still scores acceptably on the task — not the cheapest model on the list. The capability matrix sorts on any of those axes, and the cost calculator will tell you what a slot actually costs at your volume before you commit to it.

When one model is the right answer

If you work entirely through a chat interface on a consumer subscription, a stack buys you very little — you cannot route between providers inside someone else’s app, and paying for two subscriptions to save a few minutes is a bad trade. Stacks pay off on the API, in editors that let you switch models per request, and in any pipeline where you control which model handles which step. If none of those describe you, pick the best all-rounder for your main task and revisit it when your volume changes.

Your picks below save to this browser and travel in the URL, so a stack is shareable without an account. Nothing is stored on our side.