How-to · codex
How to make your Codex usage last longer
Reduce wasted Codex usage by narrowing context, limiting source material, matching output size, and choosing a smaller model.
Last verified: September 15, 2026
Direct answer
The most useful changes are to remove irrelevant context and define the task, source scope, and output length precisely. Use a smaller model for routine work. OpenAI also recommends keeping AGENTS.md focused and enabling only the MCP servers needed for the current task.
Applies to
For individuals and small teams trying to reduce avoidable consumption within a Codex plan.
What to do
- 1Separate required work from optional improvements and define the acceptance result.
- 2Provide only relevant files, directories, and date ranges instead of injecting an entire repository.
- 3Specify the audience, format, and length so the agent does not produce an unnecessary long report.
- 4Use a smaller model for routine tasks and reserve higher-capability models for difficult work.
- 5Keep AGENTS.md focused and disable MCP servers that the current task does not need.
Common misconceptions
- Longer prompts always save usage.
- Every task needs the most capable model available.
Next action
For your next task, constrain the objective, files, and output, then compare the change in your usage dashboard.
Open the card studio →