My AI setup

How I route AI work, and the tools I use

In two months my agents used about 30 billion tokens. At API list prices that is about $5,500. I paid under $200, in regular monthly subscriptions.

Most of the difference comes from one habit: the job decides the model. Below are the rules I use for that, and the tools behind them.

Last updated: 20 October 2026

The rules

  1. Plan, review, delete data, change the database: strong model. Claude Opus in Claude Code, or GPT Sol in Codex.
  2. Renames, tests, the same edit in many files: cheap model. DeepSeek V4 Flash, Gemini Flash, GLM.
  3. Risky change: a model from a different company reviews it. Models from the same company tend to make the same mistakes. When Codex leads, Claude Code reviews, and the other way round.
  4. Not sure? Ask what breaks if it is wrong. If the answer is easy to undo, the cheap model gets it.

The tools

  • Claude Codeclaude.com

    Anthropic's coding agent. One of my two leads: plans the work and reviews it, with Opus.

  • Codexopenai.com

    OpenAI's coding agent. The other lead, with GPT Sol. It reviews the risky changes Claude makes.

  • OpenCode Goopencode.ai

    A cheap monthly plan for open models like DeepSeek and GLM. Routine work goes here.

  • Command Codecommandcode.ai

    A coding agent I run on cheap models for routine work: renames, tests, bulk edits.

  • Cursorcursor.com

    An editor with AI built in. I use it to get a third look from Grok on some risky changes.

  • OpenCodexgithub.com, open source

    A router. Everything except Claude goes through it, so I see usage and cost per model in one place.

  • OpenRouteropenrouter.ai

    Pay as you go, many models through one API. No monthly plan needed.

Before you copy this

Most of these tools also have team plans, so the same setup works for a whole team, not only for one person.

Prices and limits change every couple of weeks, so check before you pay. The list above is what I use today, not a ranking.

The $5,500 is a list-price equivalent, not a bill. Most of the tokens were cache reads, which are cheap. And the Claude part is a lower bound.

No affiliate links on this page. I get nothing if you sign up.