How I Get Infinite Codex Tokens (The Goals Loophole)

Overview

The video describes a method to effectively obtain “infinite” usage of OpenAI Codex tokens by exploiting how Codex handles token limits when tasks are organized as goals.

How Codex token limits work (as described)

• Users pay a monthly subscription and receive a fixed number of tokens.
• Tokens are consumed over time, with usage refreshed on schedules (e.g., every 5 hours, every week, etc.).
• Unlike other tools (specifically compared to Claude Code), Codex does not cut you off immediately when tokens run out:
  • Claude Code: tends to stop once tokens run out (cuts off mid-work).
  • OpenAI Codex: allows the current job to finish, and then can continue running via goals.

The “goals loophole” technique

• The creator sets up long-running Codex goals in the interface.
• Example given: goals to complete all open GitHub issues.
• Behavior described:
  • Codex agents keep running until the goal is completed.
  • If the goal is framed in a way that never truly ends, the agent keeps working and keeps replenishing itself by re-queuing after finishing one unit of work.
• Key claim:
  • If the goal is “close all open GitHub issues” and the user continually ensures there are issues remaining (or keeps modifying them), the goal effectively never ends, resulting in unlimited token burn.

How the user “steers” the agent

• The only required interaction is editing/managing GitHub issues:
  • Create or adjust issues
  • Update issue content/direction so the agent keeps pursuing changing targets
• The user can re-prompt or steer further, but the core control mechanism is issue updates, which guide the agent’s direction without heavy micromanagement.
• Mentioned constraint: pressing Escape stops pursuit of the goal, which would end the run.

Practical implication / use case example

• The creator mentions a real workflow: porting a backend to a new database where an agent could run for multiple days (e.g., ~5 days) under a goal setup.
• They emphasize the method is “not that much to micromanage” because steering is primarily done through GitHub issue management.

Key conditions highlighted

• “Infinite tokens” is claimed to work as long as the goals don’t end.
• Using GitHub issues as a continuously maintained work queue is presented as the method to keep goals active.

Main speakers/sources

• The video creator/speaker (no other named sources in the subtitles).
• Indirect comparisons mentioned: OpenAI Codex and Claude Code (Claude by Anthropic), referenced as tooling behaviors.