A common issue I started facing when using a coding agent such as Claude Code or Codex was that my usage was going extremely fast. By extremely fast, I mean that for Codex, I would typically use up my weekly allowance in about 10-12 hours, which becomes very expensive if I am to cover everything for a week with my Codex subscriptions.

I started thinking about what causes my subscriptions to go that fast, and I think there are numerous reasons for this, for example: My code repository gets larger over time. We get access to better models, such as Claude Fable or GPT-6-Astra, which cost more to run and thus increase the usage faster. The companies are less generous with the subscription coding plans, as they cost them a lot of money to run. We simply code more, and more agents in parallel.

Thus, I had to figure out a way to make my coding and subscription plans last longer while impacting quality minimally. This article will be a summary of my experiences with coding agents and how I get more out of my subscriptions, and the specific techniques that I apply every day when I code.

On a high level, I use the following techniques: Use smaller models for simpler tasks (GPT-5.6 SOL, Claude Opus). Use the smartest models (Fable and GPT-6) for orchestration, and make sure they spin up subagents with smaller models. Refactor the codebase regularly (complicated codebase, requires more reasoning from the models, costing more tokens, etc.). Avoid god files. Clean up your markdown files (AGENTS.md, CLAUDE.md, skills, hooks, etc).