ChatGPT Plus Codex Limits: Messages, 5-Hour Window & Weekly Cap

👁 116 reads

In ChatGPT Plus, Codex doesn’t have a fixed number of tasks. OpenAI gives you ranges per 5-hour window: from 5-45 messages in GPT-6 Astra to 350-3000 in GPT-6 Luna, and a weekly limit may also apply. Below, we explain what affects usage, how to check it, and what to do when you hit the limit. Everything about the agent itself is in our Codex from OpenAI guide.

How Much Codex You Get with ChatGPT Plus

These are official estimates from OpenAI’s pricing page for messages executed locally (desktop app, CLI, IDE extension) in one 5-hour window:

OpenAI itself notes that these are “not fixed message limits.” The 5-45 range exists because one message to the agent can mean reading three files or an entire repository, running tests multiple times, and lengthy reasoning. Similar-looking tasks consume different amounts because what counts is the model, context, reasoning level, tools, and cache, not the length of your instruction.

5-Hour Window and Weekly Limit

• 5-hour window applies to Plus, Business Standard, and Edu and Enterprise without flexible billing (these have per-person limits like Plus).
• Weekly limit - the pricing page only says it “may also apply.” OpenAI doesn’t publish the exact value. The usage panel shows the reset time for both limits.
• Pro 100, Pro 200, Pro 500 - no 5-hour limit, but with a total allocation depending on the plan.
• Enterprise and Edu with flexible billing - no fixed limits, they pay with credits.

If you hit the limit mid-task, the agent will finish the current turn (subject to “fair use”). It won’t start a new one until the limit resets or you buy credits.

What Burns Through Your Limit Faster

• Cloud instead of local - tasks in Codex in the cloud can consume more than local ones, according to OpenAI.
• Fast mode - 2.5 times faster subscription limit consumption. Astra Ultrafast - 8 times, but this mode is only in Pro 500.
• Image generation - on average 3-5 times faster than a similar turn without images. The Free plan doesn’t include images in Codex.
• Voice in desktop app - 0.05 USD per minute from your Codex budget, plus the cost of the model executing the task.
• MCP servers and long AGENTS.md - each adds context to every message.
• Partner apps - Devin, Notion, Vercel, and other tools connected via Sign in with ChatGPT draw from the same pool. In Plus and Pro, this also applies to ChatGPT for Excel.

Only code review triggered by GitHub (marked @codex review or automatic reviews) counts separately. Local review with the /review command uses the general limit.

Models Consume Differently: Calculation for One Task

OpenAI publishes rates in credits per million tokens. These don’t tell you how many messages fit in your subscription, but they show the proportions between models well. We calculated one example task using these rates: the agent reads 200,000 tokens of context (150,000 from cache) and writes 20,000 tokens of output.

Conclusion: one message in Astra costs about as much as five in GPT-6.1 Sol and about one hundred in Luna. This matches the ranges from the table above (Astra 5-45, Sol 15-160). At the starting credit price OpenAI announced in October 2025 (1000 credits for 40 USD), this task would cost about 1.65 USD in Astra, 0.32 USD in GPT-6.1 Sol, and 0.02 USD in Luna. Today OpenAI says credit pricing depends on your plan, so check yours in settings before purchasing.

ChatGPT Plus Compared to Other Plans

Business plans in USD according to OpenAI’s pricing page:

Current OpenAI pages don’t specify exact multiples for Pro 100 and Pro 200 (at Pro 100’s launch in April 2026, it was said to be 5 times more than Plus). Anyone who had active Pro 200 between September 22 and 29, 2026, keeps their previous allocation until October 29, 2026, then gets a lower one at the same price. Full comparison of plans is in our ChatGPT pricing article, and we cover the most expensive plan in ChatGPT Pro 500.

How to Check Your Usage

1. In CLI: during a session, type /status. It shows chat ID, context occupancy, and limits. We describe CLI installation in Codex CLI on Windows.
2. Usage panel: chatgpt.com/codex/settings/usage - current limits and reset times.
3. In ChatGPT: Settings → Usage, and in the app Codex Usage & Billing. You can also buy credits there and set limits for partner apps (App limits).

OpenAI recommends checking the panel every week or two. We covered general chat limits in ChatGPT limits.

What to Do When You Hit the Limit

• Wait for reset - ChatGPT shows the reset time. OpenAI support won’t manually reset your limit.
• Switch to another model - according to OpenAI help, the model that hit the limit may be unavailable until reset, but you can work with another available one in the meantime, e.g., switch from Astra to Sol.
• Buy credits (Plus and Pro) - your plan limit is used first, then credits. You don’t need to change plans.
• Switch to API key - additional local tasks at API rates (e.g., GPT-6.1 Sol 2 USD / 10 USD per million input and output tokens). Without cloud features.
• Upgrade to Pro - only makes sense if you regularly hit the limit, not once a month.

How to Save Your Limit

• Work on GPT-6.1 Sol by default, and switch to Astra with the /model command only for difficult tasks. Set your default model in config.toml with the line model = "...".
• Give routine work (renaming, formatting, simple scripts) to Luna.
• Don’t enable Fast mode if you’re not waiting for results.
• Point to specific files instead of “review the project,” and shorten long sessions with the /compact command.
• Shorten AGENTS.md and disable unused MCP servers - these are official recommendations from the pricing page.
• Set weekly caps for partner apps in App limits, e.g., 10-20% to start.
• Work locally, and save the cloud for tasks that should run with your laptop closed.

How Much My Test Consumed

We tested this on a real task. On October 3, 2026, Codex CLI 0.160.0 received an instruction to build a single-file gross-to-net salary calculator for 2026 (12-month table, tax threshold, 30x limit). We tested it in Polish.

• codex exec mode only shows a token counter, not a percentage of your limit. Check the percentage with /status in an interactive session or on the usage page.
• On Astra, this task is expensive in terms of limit. With the same number of tokens, GPT-6.1 Sol would have consumed roughly 5 times less according to the rates in the table above.

Full test description with control values and comparison with Kimi Code and a local model is in Same gross-to-net calculator in Codex, Kimi Code, and locally.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and Verification Date

• OpenAI: Codex / ChatGPT Work pricing and limits (ranges, Fast and Ultrafast multipliers, credit rates, tips)
• OpenAI Help Center: Pro plans (Pro 200 until Oct 29, 2026, credits, limit resets)
• OpenAI: session commands (/status)
• OpenAI Developers on X (credits: 1000 for 40 USD, October 2025)
• OpenAI Forum: Pro 100 plan (April 2026)

Business plans and API are in USD. The credit calculation for the task is our own math using official rates, not a measurement. Verified October 3, 2026.