1. Compress the context before the model sees it
The single biggest source of spend is bulky, repetitive context: build logs, test output, JSON fixtures, and file content replayed turn after turn. Headroom sits between your machine and the API and reversibly compresses that noise locally: about 50% fewer tokens on the noisy items it compresses, typically 40-50% across a session, with the original content retrievable whenever the model needs the detail. One install covers the ChatGPT app, the CLI, and the IDE extension, because all three read the same ~/.codex/config.toml.
2. Make the agent write less
Output tokens drain the same windows as input. Ponytail is a skill that nudges the agent to write the least code that solves the problem: smaller diffs, fewer output tokens, less to review. Under token metering, verbose answers are charged directly, so this converts straight into more work per window.
3. Keep noisy output out of the transcript
A failing build can dump thousands of lines of stack traces and dependency noise into the conversation, and every later turn replays it. Pipe test runs through quieter reporters, keep generated files out of what the agent reads, and convert heavy documents to plain Markdown before attaching them. The Codex cost guide walks through the tools for each of these, and everything in it applies unchanged after the ChatGPT merger.
4. Watch the windows instead of upgrading
When you hit a limit, OpenAI offers credits or a bigger plan. Before paying, look at where the tokens went: the ChatGPT coding usage limits page explains the 5-hour and weekly windows, and Headroom's menu bar shows both live so you can pace heavy sessions instead of discovering the wall mid-task.