How to check your Codex usage, and what it costs in tokens

If you are reading this you have probably just hit a usage-limit notice and want to know how much is left and when it resets. The first half of this post answers that from OpenAI’s documentation. The second half covers what OpenAI’s meter leaves out, which is what a Codex session actually spends in tokens, energy and carbon.

Where to look

OpenAI shows the allowance in three places. In ChatGPT, Settings and then Usage shows the allowance your plan includes and any credit balance. In the Codex app, the same view is under Usage and billing. And when you hit a limit, the notice in Codex links straight to it. Members of workspaces on credit-based billing see a monthly Work and Codex allowance and their credit history under Settings and then Usage and billing in ChatGPT Desktop. Practitioners also report that the Codex command line can show the meters on its status line, which saves opening the page.

How the limits work

Codex has no limit of its own; it draws from the agentic usage allowance of your ChatGPT plan, shared with the other agentic features, and the allowance is metered in two units: local messages, for work on your machine, and cloud tasks, for work on OpenAI’s infrastructure. Two limits apply at once, a rolling five-hour window that spreads use across the day and a separate weekly cap above it, so you can exhaust the window with most of the week still in hand. Your plan’s included usage is spent first; after that, usage draws from credits you buy, which top up any supported feature on the account.

The size of the allowance depends on the plan and the model, the Pro tiers carry multiples of the Plus allowance, and OpenAI changes the numbers. In August 2026 it restored the five-hour window for Plus subscribers after a period without one, while leaving the Pro tiers exempt for the following months. Any table of message counts is out of date within weeks, so this post does not reproduce one; OpenAI’s help page has the current version.

What the meter does not tell you

A message or a task is not a fixed amount of work. A Codex task that reads a handful of files and edits one is a few thousand tokens; a task that reads a large repository and rewrites a module is tens of thousands. Two tasks count the same against your allowance and cost very different amounts of compute. And nothing in OpenAI’s meter says what any of it used in energy or emitted in carbon, which is the number a company reporting AI usage needs.

Counting the tokens yourself

The crbon CLI hooks Codex on your machine, alongside Claude Code, Gemini CLI, Cursor and OpenClaw, and records each session’s tokens per model from the agent’s own logs. It syncs on a schedule, and the dashboard shows the sessions per agent, per model and per week, with the cost, the energy and the CO₂e computed under methodology v1.2. Redaction runs before anything leaves the machine, and the metering needs only metadata. Your plan’s limits stay in OpenAI’s app; the CLI reports what you used, not what you have left.

What a Codex turn costs

Codex runs on OpenAI’s reasoning models, so the reasoning tier is the closer fit; the medium tier is shown for comparison. The figures are per turn, under methodology v1.2, with on-site water at Berkeley Lab’s 2023 US average of 0.36 litres per kilowatt-hour.

For scale, a chat query on the same factors is about 0.18 watt-hours, so a large Codex turn on a reasoning model is forty times a chat message. A developer running a hundred such turns a day spends about 0.8 kilowatt-hours and emits about 365 grams of CO₂e, a third of a kilogram, before the water behind the electricity is counted. Multiply that across a team and it becomes the figure a company is now being asked to report.

Keeping it down

The allowance and the footprint move together, because both are spent in tokens. Smaller contexts help most: point Codex at the files that matter rather than the whole repository, and split a large change into tasks that each read less. Reuse a session where the context is already loaded instead of starting cold. And use the reasoning-heavy modes for the work that needs them; a large turn on the reasoning tier costs twice a large turn on the medium tier for the same tokens. The energy calculator in our post on how much energy AI uses runs the same arithmetic for any turn size.