CODEX USAGE DASHBOARD + CYCLE ANALYZER
OpenAI's dashboard tells you where your usage went. It does not tell you whether you are going to run out. This page explains every block of the real dashboard, then analyzes your own burn across the current weekly cycle — in your browser, without touching your account.
Open your dashboardRun /status in the Codex CLI and paste the whole thing. One paste fills every window it reports — percentage, reset time, countdown.
Paste your /status above, or enter the weekly percentage by hand.
Read the weekly meter off the dashboard and enter it here.
Everything here stays in this browser. No account, no upload, no cookie.
Call it. You will see what everyone else picked once enough forecasts are in — and every poll is scored against what actually happened.
What this site's model says:
These percentages are reader expectations, not this site's forecast. The model number beside them is calculated from verified reset history and is never mixed with the vote.
Track record: No poll has resolved yet. The first scored result appears after the next verified reset.
| /status in the CLI | the analytics dashboard | |
|---|---|---|
| answers | how much is left right now | where it went over the last days |
| 5-hour window | yes — live percentage and reset time | not shown at all |
| weekly window | yes — live percentage and reset time | yes — percentage and reset time |
| history | no | yes — 7 days, 1 month, or a custom range |
| by model / surface / skill | no | yes |
| freshness | read at the moment you run it | aggregated, lags by minutes |
Run /status when you are deciding whether to start another task. Open the dashboard when you want to know which model, surface or skill consumed the window.
Observed directly on the live page rather than copied from documentation. OpenAI's own enterprise guidance warns against treating dashboard labels as a stable contract, so this section carries its own observation date.
The top card. It states plainly that Codex and Work share the same usage limit, then shows your weekly usage limit as a percentage remaining with an absolute reset timestamp — the same server-sent time the CLI renders.
easy to misread: It is a percentage, not a quantity. OpenAI publishes no weekly cap for any plan, so 20% remaining cannot be converted into a number of messages.
Underneath the weekly bar sits a separate GPT-5.3-Codex-Spark meter with its own percentage and its own reset time — a different clock from the weekly window, not a breakdown of it.
easy to misread: Draining Spark does not drain your weekly window, and a weekly reset does not refill Spark. Two meters, two independent rollovers.
Credits extend usage past the plan limits and are spent only after a window is exhausted. Auto-reload sits directly below it and tops the balance up automatically when it runs low.
easy to misread: Zero credits is the normal state for someone who has never bought any — it is not an error, and it is not your weekly limit hitting zero.
A 0–100% chart across the selected range, split into Desktop App, CLI, Web and Exec. This is the block that tells you which client is actually eating the window.
easy to misread: Surfaces are clients, not budgets. All four draw on the one shared pool, so a quiet CLI does not mean you have headroom.
The total number of turns in the range, toggled between By model and By surface. The model list is the live routing set — Sol, Terra and Luna alongside the older 5.5, 5.4 and 5.4-mini, plus Spark and the automatic code-review model.
easy to misread: A turn is not a unit of cost. A long Sol turn at high reasoning effort and a one-line Luna turn both count as one.
Two counters over the same range: how many plugin calls ran, and how many skills were used, broken down per skill.
easy to misread: These count invocations, not consumption. They are useful for spotting a skill that fires far more often than you expected, not for attributing budget.
A paged table of Date, Service and Credits used. It stays empty until credits are actually spent.
easy to misread: It records credit spend only. Ordinary in-plan usage never appears here, so an empty table says nothing about how heavy your week was.
Range is 7D, 1M or a custom window, with a Group by selector that buckets the charts (Day by default). A second tab, Code review, covers automatic reviews separately from everything above.
The range applies to the history charts only. The balance card at the top is always live and ignores it.
Dashboard anatomy observed 2026-08-02
The dashboard shows one number today. Log it a few times across the week and the shape appears: how fast you are actually burning, how much you can spend per day and still make it, and when you run out if nothing changes.
sustainable pace your burn
One real account, sampled every minute by this site's reset verifier and published as a daily curve. It is a single account on a single plan running one person's workload — a shape to compare against, not a population estimate.
Source: this site's own monitored account. Not OpenAI data, and not an average of users.
It is inside Codex settings, not ChatGPT settings, under Analytics — the page is titled "Codex and Work Analytics" and the Usage tab is the default. The older /codex/cloud/settings/usage address is no longer the surface that carries the meters. https://chatgpt.com/codex/cloud/settings/analytics#usage
Run /status in an active Codex CLI session. It prints the live percentage remaining and the reset time for each window, which is the one thing the dashboard does not give you for the 5-hour window.
It never has. The analytics page is built around history and attribution over a date range; the 5-hour window is a live rolling budget with nothing meaningful to plot. Only /status and the clients themselves report it.
A separate allowance for the Spark model with its own percentage and its own reset timestamp, shown directly beneath the weekly bar. It runs on an independent clock — spending it does not touch your weekly window, and the weekly rollover does not refill it.
Yes. The balance card says so outright: Codex and Work share the same usage limit. ChatGPT for Excel and Workspace Agents draw on that same agentic pool. Ordinary ChatGPT chat is metered separately and never drains it.
Unknowable from the dashboard. OpenAI publishes per-5-hour message ranges per plan but no weekly cap for any plan, so the weekly meter is a percentage of an undisclosed total. Anyone quoting an exact weekly message count is guessing.
The history charts aggregate and lag by minutes, and the surfaces share one pool — a cloud task, an automatic code review or a Work session can consume budget while your terminal sits idle. Compare against /status, which reads live.