Cost Optimization Guide
How to combine tools to maximize quality and minimize cost — savings of 60-70% on average.
Cost Architecture
Claude Code
Claude Code — Sonnet/Opus via subscription. Front-end, UI, creativity, architectural decisions.
Codex
Codex — GPT-5-Codex via subscription. Code review, adversarial review, background rescue.
OpenRouter
OpenRouter (Qwen/GLM) or Ollama (local). Deterministic tasks, boilerplate, refactoring.
Principles
- 1
Claude Code is the orchestrator — use where it is irreplaceable: UI, UX, creativity, complex decisions
- 2
Codex is the reviewer — use for code review, adversarial review, and task delegation
- 3
OpenRouter is the cheap executor — use Qwen 3.6 or GLM 5.1 for deterministic tasks
- 4
OpenCode + Ollama is the free executor — use when you want absolute zero cost
Step-by-Step Setup
1. OpenRouter as cheap provider
Get API key at openrouter.ai and configure in .codex/config.toml with model_providers.openrouter.
[model_providers.openrouter]
name = "OpenRouter"
base_url = "https://openrouter.ai/api/v1"
env_key = "OPENROUTER_API_KEY"
wire_api = "responses"2. Custom agent with cheap model
Create .codex/agents/qwen-worker.toml with model = "qwen/qwen-3.6-plus" and model_provider = "openrouter".
name = "qwen-worker"
description = "Executor de baixo custo para tarefas determinísticas."
model = "qwen/qwen-3.6-plus"
model_provider = "openrouter"
sandbox_mode = "workspace-write"