What changed
Under message-count metering, a one-line question and a monster debugging turn cost the same: one message. Under token metering, every prompt, file read, tool call, and model response is weighed — a heavy request drains the 5-hour window proportionally faster. The weekly cap on top works the same way.
When you do run out, OpenAI gives you two levers: buy additional credits, or drop to a smaller model until the window resets. One costs money, the other costs capability.
Who this actually hurts
Not the person asking Codex short questions — their turns were always light. It hurts the workflows where most tokens are machine-generated: debugging from long logs, codebase exploration that reads dozens of files, shell-heavy tasks, and anything that loops over a failing build. Under message metering those turns were subsidized; under token metering you pay for exactly what the model reads, and the model reads far more than you type.
The upside nobody mentions
Message metering gave you no way to stretch your plan — a message was a message. Token metering does: if your requests get lighter, your plan gets bigger. Efficiency now pays directly. Sending only the failing part of a build log instead of all of it, starting fresh sessions for new tasks, keeping generated files out of exploration — each one now buys you real headroom in the window.
Making it automatic
Doing that manually is a chore, which is exactly why we built for it. Headroom routes the Codex CLI through a local proxy and reversibly compresses tool output, logs, and boilerplate before they reach the model — around 50% fewer tokens per request, which under token metering translates directly into roughly twice the usage on the same ChatGPT plan. It also tracks your Codex usage windows from the menu bar, so the cap stops being a surprise.
Related reading: how Codex usage limits work, what burns Codex usage fastest, and the full Codex cost-cutting toolkit.