The $0.02/Call Coworker Pattern That Solves the Budget Problem

A Reddit post about giving Claude Code a cheap coworker hit 1,123 upvotes and 115 comments on r/ClaudeAI. Read together with the Uber story above, this is the demand signal paired with its solution.

The setup: route routine implementation work to a $0.02/call model (Gemini Flash, GPT-4o-mini) while keeping Opus for architectural decisions and complex debugging. The author tried everything else first. Compact mode, Sonnet for simple tasks, tighter prompts. None of it worked. The coworker pattern did.

The insight that makes this click: 60-70% of agent tool calls are read-only operations. File reads, grep searches, directory listings, simple edits. These don't need frontier reasoning. They need fast, cheap execution. You're paying Opus prices for tasks that a $0.02 model handles identically.

Multiple implementations now exist. LiteLLM proxy routes by task type. Claude-code-proxy adds model routing middleware. Kong AI Proxy handles it at the infrastructure level. This isn't one person's hack anymore. It's a cross-tool strategy that works with Claude Code, Cursor, and Aider.

I've been thinking about this as the "tiered workforce" model for AI. You don't send a senior architect to rename variables. You don't need Opus to read a file. The same workforce management principles that work for human teams apply to model selection. Match the capability to the task.

The math works out to roughly 3-5x cost reduction on a typical Claude Code session without sacrificing output quality on the tasks that matter. If Uber's engineers were spending $1,000/month, this pattern brings it to $200-300. That's the difference between a blown budget and a sustainable line item.

What to do about it: Set up liteLLM or claude-code-proxy this weekend. Route file reads, greps, and simple edits to Gemini Flash. Keep Opus for architecture, debugging, and anything requiring multi-file reasoning. The setup takes about 30 minutes and pays for itself on day one.