Electricity Bench

Codex CLI usage limits

What the 5-hour window and the weekly limit are on OpenAI's plans, when they reset, what OpenAI publishes and what it does not, and how many real coding tasks fit in a week on each plan, from our W38 2026 run.

On ChatGPT Pro 20x, GPT-5.6 Sol fits about 2500 tasks a week through Codex CLI, estimated from the tier multiples. That is 130 passes of our fixed task set (real-world issues plus spec planning, 19 tasks) in one weekly window. A task is a whole multi-turn agent session, and the share it used is read off OpenAI's own meter.Measured

Provider pages reviewed Sep 14, 2026. Figures from the W38 2026 grading run, published Sep 14, 2026.

A 5-hour window with a weekly cap behind it

OpenAI publishes Codex limits as message estimates per five-hour period: "The estimates below show local messages per five-hour period", with 10-100 messages on Plus, 50-500 on Pro 5x and 200-2,000 on Pro 20x for GPT-5.6 Sol (learn.chatgpt.com/docs/pricing, retrieved 2026-09-14). The same page says "Local messages and cloud chats share your plan's usage allowance. Weekly limits may also apply."

The meter Codex CLI itself reads reports both windows as a percent used with a reset instant: a 300-minute window and a 10,080-minute one. Our prober takes both before and after every run, and every Codex row on the leaderboard is read on the weekly window. A message estimate is not a task estimate: one benchmark task is a multi-turn agent session, which is why the table below counts passes of a fixed task set rather than messages.

When the limits reset

Both windows are rolling. Each one reports its own reset instant on the meter, and Codex CLI shows it when you hit the limit. OpenAI publishes no fixed reset weekday. Which window the meter calls primary is OpenAI's choice and has changed once during our measurements, so the page reads the weekly window regardless of which slot it sits in.

What a tier multiplies

OpenAI sells Pro as 5x and 20x the Plus allowance (chatgpt.com/pricing, retrieved 2026-09-05), and its message estimates scale by the same factors. The table scales our ChatGPT Pro 5x measurement by those advertised multiples to fill the Plus and Pro 20x rows, marked est, the way the leaderboard's budget picker does. It never scales by price.

Reasoning effort and speed settings

The pricing page states that "Speed configurations increase credit consumption for all applicable models" and says nothing about reasoning effort (learn.chatgpt.com/docs/pricing, retrieved 2026-09-14). Every Codex subject we grade runs at model_reasoning_effort=medium, so this page states no multiple for effort high or max.

Shared with ChatGPT chat

The pricing page confirms that local Codex messages and Codex cloud chats share one allowance. Whether ordinary ChatGPT chat draws from the same weekly allowance is not stated on the pages we retrieved, so this page does not claim it either way.

How many tasks fit in a week

One suite run is one pass of our fixed task set: real-world issues plus spec planning, 19 tasks. Each cell is how many of those passes one weekly window of the plan fits, from the share one pass consumed of the provider's own meter. A task is a whole multi-turn agent session on a real codebase, so this reads far lower than a message count would. Rows marked est scale the ChatGPT Pro 5x measurement by OpenAI's advertised usage multiples (chatgpt.com/pricing, retrieved 2026-09-05), never by price; we have not run on those plans.

PlanGPT-5.6 Soloverall CGPT-6 Astraoverall CGPT-5.6 Terraoverall D+GPT-5.6 Lunaoverall D
ChatGPT Plus$20/moabout 6.7 runsestabout 130 tasksabout 3.3 runsestabout 63 tasksabout 10 runsestabout 190 tasksabout 20 runsestabout 380 tasks
ChatGPT Pro 5x$100/mo · measuredabout 33 runsabout 630 tasksabout 17 runsabout 320 tasksabout 50 runsabout 950 tasksabout 100 runsabout 1900 tasks
ChatGPT Pro 20x$200/moabout 130 runsestabout 2500 tasksabout 67 runsestabout 1300 tasksabout 200 runsestabout 3800 tasksabout 400 runsestabout 7600 tasks

Hover or focus a cell for the derivation. "More than" marks a run the meter could not resolve: it read under one percent for the whole pass, so the count is a floor. A share is always of the plan's own allowance and is not comparable across providers.

The plans

ChatGPT Plus at $20/month: not run on; its row is an estimate from the measured plan. Is ChatGPT Plus worth it? · scorecards: GPT-5.6 Sol, GPT-6 Astra, GPT-5.6 Terra, GPT-5.6 Luna

ChatGPT Pro 5x at $100/month: the plan our subjects grade on, so its row is measured. Is ChatGPT Pro 5x worth it? · scorecards: GPT-5.6 Sol, GPT-6 Astra, GPT-5.6 Terra, GPT-5.6 Luna

ChatGPT Pro 20x at $200/month: not run on; its row is an estimate from the measured plan. Is ChatGPT Pro 20x worth it? · scorecards: GPT-5.6 Sol, GPT-6 Astra, GPT-5.6 Terra, GPT-5.6 Luna

What OpenAI publishes

FigureStatus
Local window length (5 hours)published
Message estimates per 5-hour window, per planpublished
Tier multiples (Plus 1x, Pro 5x, Pro 20x)published
That a weekly limit existspublished
Size of the weekly limitnot published
Reset weekday or hournot published
Reasoning effort multiplenot published
Whether ChatGPT chat shares the Codex allowancenot published
Percent of each window one run consumedour meter reads it

Sources: Codex pricing (learn.chatgpt.com/docs/pricing), retrieved 2026-09-14; chatgpt.com/pricing, retrieved 2026-09-05. Our own meter readings are described on each scorecard under "How the plan cost was measured".

Frequently asked questions

How many Codex CLI tasks a week does ChatGPT Plus give you?

In one weekly window of ChatGPT Plus, one pass of our fixed task set (real-world issues plus spec planning, 19 tasks) fits GPT-5.6 Sol: about 6.7 passes (about 130 tasks), estimated; GPT-6 Astra: about 3.3 passes (about 63 tasks), estimated; GPT-5.6 Terra: about 10 passes (about 190 tasks), estimated; GPT-5.6 Luna: about 20 passes (about 380 tasks), estimated. Figures are estimated from our ChatGPT Pro 5x measurement by OpenAI's advertised tier multiples, W38 2026. A benchmark task is a whole multi-turn agent session, not a message.

How much of ChatGPT Pro 5x does one Codex CLI task use?

About 0.2% of the weekly window on GPT-5.6 Sol: one weekly window of ChatGPT Pro 5x fits about 630 tasks, read off the provider's own meter before and after each graded run, W38 2026. A benchmark task is a whole multi-turn agent session, not a message, so a provider's message estimate is a different unit.

When does the Codex weekly limit reset?

Seven days after the window opened, on a rolling basis. The meter Codex CLI reads reports a reset instant for the 5-hour window and for the weekly window, and the CLI shows it when you hit a limit. OpenAI publishes no fixed weekday (learn.chatgpt.com/docs/pricing, retrieved 2026-09-14).

Does reasoning effort high burn the Codex limit faster?

OpenAI publishes no figure for it. Its pricing page says speed configurations increase credit consumption and is silent on reasoning effort. Every Codex model we grade runs at effort medium, so we have no same-model comparison to quote.

Is the Codex limit shared with ChatGPT chat?

Local Codex messages and Codex cloud chats share one allowance, per OpenAI's pricing page (retrieved 2026-09-14). Whether regular ChatGPT chat draws from the same weekly limit is not stated on the pages we retrieved.

Other harnesses: Claude Code · Antigravity CLI. The leaderboard shows the same usage per plan-price band; compare subscriptions puts the plans side by side.