Claude and Codex Usage Limits in 2026: 5-Hour Windows, Weekly Caps and 7 Ways Round Them (Claude, ChatGPT, Codex, Jobbit)

It is 11:40 on a Tuesday, the build is half finished, and the message appears: "You have reached your usage limit. Resets at 16:00." Anyone who pays for Claude or ChatGPT in 2026 knows the feeling, and the searches show it: "claude usage limit", "claude 5 hour limit reset", "codex usage limits" and "chatgpt pro waitlist" have all climbed this year as the two most popular AI subscriptions moved from unlimited chat to metered windows. The tools are better than ever; the meter is the new problem.

This guide explains the Claude usage limits and the Codex limits as they stand in September 2026: the rolling five-hour window, the fixed weekly cap that sits on top of it, what each plan actually buys, why OpenAI stopped selling its $200 Pro plan to new customers, and the working habits that stop a limit landing in the middle of a job. It ends with the honest alternative for people who cannot plan their day around a reset timer: an agent that meters by a monthly credit pool instead of by the clock.

How we checked the limits

1. Vendor documentation first. Anthropic's and OpenAI's own help pages for plan limits, session windows and reset behaviour, read in September 2026.

2. Dated changes, not folklore. Both companies changed limits several times this year; we list what applies now and when it changed.

3. Ranges, not promises. Neither vendor publishes a fixed message count. Usage depends on model, message length, attachments and conversation length, so we quote their published ranges and say so.

4. Real working days. The practical section comes from running client work through Claude Code, Codex and Jobbit for a month and noting where each one stopped.

Claude, ChatGPT and Codex limits at a glance

Prices are in the currency the vendor lists. Claude and Codex figures are the vendors' published ranges for September 2026 and move with the model and the length of the conversation.

How the Claude 5-hour limit actually works

Claude does not count messages. It counts usage inside a rolling five-hour session that starts with your first message and ends five hours later, whatever you do in between. Pro buys "at least five times" the Free allowance per session, Max 5x five times Pro, Max 20x twenty times Pro. A long conversation with big files attached burns the window far faster than short questions, because every reply re-reads the whole thread.

Since 2025 there is a second meter: a weekly cap with a fixed reset time assigned to your account, which is the one people hit after three or four heavy days in a row. The five-hour window is the limit you feel on an ordinary day; the weekly cap is the one that ends a busy week on Thursday.

Claude Code, the terminal agent, shares the same allowance and consumes it quickly, which is why Anthropic doubled its five-hour limits in May 2026 and, after a summer promotion of 50% extra weekly Code usage, made a permanent 25% weekly increase from 14 September 2026. Max 20x users can also buy additional usage. None of this changes the shape of the problem: a window, a cap, and a reset time you plan around.

How the Codex and ChatGPT Pro limits work

Codex, OpenAI's coding agent, meters by rolling five-hour windows per model, shared across every surface (CLI, IDE extension, web and the iOS app) on one plan allowance, plus a weekly cap that applies after several heavy days. On Plus with the GPT-5.6 family the published ranges are roughly 10 to 100 messages per five hours on the largest model (Sol), 25 to 200 on the middle one (Terra) and 250 to 2,000 on the fast one (Luna). There is no fixed reset clock; capacity frees up as usage older than five hours ages out. OpenAI tightened these ranges earlier in 2026, restored them after complaints, and states plainly that they can change again.

The Pro tiers multiply the allowance: the $100 plan is 5x Plus, the $200 plan is 20x Plus. Then, on 10 September 2026, OpenAI paused new sign-ups for the $200 Pro plan worldwide, citing "unprecedented" demand for GPT-6 Astra and its computer-use feature, which consumes capacity faster than earlier models. Existing Pro subscribers keep their access; new customers cannot buy the top tier and no reopening date has been given. Plus, Go, Business, Enterprise and the API stayed open. For anyone who was planning to pay their way out of the limits, the door is currently shut.

What a limit costs in a working day

• The build that stops at 80%. Claude Code or Codex runs a refactor for 40 minutes, hits the window, and the half-applied change sits in your repository until the reset.

• The client call you cannot prepare for. The research you needed at 2pm is available at 4pm.

• The team that shares one login. Business plans meter per seat; a shared account hits the cap by lunchtime.

• The "just try the other tool" tax. Pasting context from Claude into ChatGPT, or the reverse, costs 20 minutes each time and produces two half-informed assistants.

• The upgrade that is not for sale. The $200 plan that removed most of the pain is paused for new customers.

7 ways round Claude and Codex limits

1. Start the clock deliberately. The five-hour window opens with your first message. Open it when you start the real work, not with a throwaway question at 9am that leaves you dry at 1pm.

2. Keep threads short. Each reply re-reads the whole conversation. Start a new chat for a new task and paste a three-line summary; it can halve your consumption.

3. Use the cheapest model that can do the job. Sonnet-class or the fast Codex model for drafting, edits and boilerplate; the big models only for the hard step. On Codex the fast model allows roughly ten times the messages of the largest one.

4. Attach less. One relevant file beats the whole folder. For code, point the agent at the directory rather than pasting it.

5. Batch the heavy work. Group the three big jobs into one session late in the day, let the window expire overnight, and use the morning for cheap tasks.

6. Run two providers, not one account twice. A Claude plan for reasoning and a ChatGPT plan for Codex gives two independent windows for $40 a month and a second opinion for free.

7. Move the metered work to an agent that meters differently. The limits above exist because chat subscriptions sell time-boxed capacity. An agent that sells a monthly pool of credits lets you spend a quiet Sunday's allowance on a frantic Tuesday.

An agent with no hourly, daily or weekly windows

ChatGPT, Claude, Gemini, Microsoft Copilot, Manus and Meta's Muse all cap you by the clock in one way or another: a session window, a daily quota, a weekly ceiling. Jobbit meters by work done instead. Paid plans come with a monthly credit pool (6,000 credits on Plus at £15 a month, 35,000 on Pro at £79, 80,000 on Max at £159) and no five-hour, daily or weekly windows: if the launch is on Tuesday, you can spend half the month's credits on Tuesday. When the pool runs out you top it up with a credit pack rather than waiting for a reset. The free plan is the exception, with 100 credits a day, which is how you try it.

Three other things change the arithmetic:

• Model choice inside one job. Jobbit's model picker lets you run Claude for the reasoning step, GPT for the next and a cheap model for the boilerplate, so you are not paying big-model credits for small work.

• Concurrent tasks instead of one window. Plus and Pro run 10 tasks at once, Max runs 20, so the refactor, the research and the invoice run side by side rather than queueing inside one session.

• The job finishes, not just the chat. The agent builds and hosts the app, writes the file, runs the automation, as described in how to build and deploy a web app with AI, which means fewer round trips and fewer credits per outcome.

Credits are still finite, and long autonomous runs on the strongest models are the expensive ones; the full breakdown is in Jobbit pricing explained. The difference is that the limit is a number you control, not a timer you obey. Our ChatGPT alternatives guide compares the wider field.

Mistakes to avoid

• Opening the window with small talk. The five hours start with the first message; make it count.

• Letting one thread run all day. Long threads are the fastest way to burn a Claude session.

• Assuming the limit is a message count. It is usage; a 200-page PDF costs more than 50 short questions.

• Sharing one account across a team. Buy seats or move to a plan that meters per month.

• Waiting for the $200 Pro plan. Sign-ups are paused with no date; plan for the limits you have.

• Paying for capacity you never use. If you only need heavy usage on launch weeks, a monthly pool suits you better than a weekly cap.

Frequently asked questions

What is the Claude 5-hour limit?

Claude meters usage inside a rolling five-hour session that starts with your first message. Pro gives at least five times the Free allowance per session, Max 5x five times Pro, and Max 20x twenty times Pro. When the allowance is used you wait for the window to roll on; paid plans also have a fixed weekly cap on top.

When does the Claude usage limit reset?

The five-hour session resets on a rolling basis, five hours after it started, and the app shows the exact time. The weekly cap resets at a fixed time assigned to your account, shown in usage settings. Claude Code shares both allowances, with a permanent 25% weekly increase for Code from 14 September 2026.

What are the Codex usage limits on ChatGPT Plus?

Codex uses rolling five-hour windows per model, shared across the CLI, IDE, web and iOS. OpenAI's published Plus ranges for the GPT-5.6 family are about 10 to 100 messages per five hours on the largest model, 25 to 200 on the mid model and 250 to 2,000 on the fast model, plus a weekly cap. The $100 Pro plan is 5x and the $200 plan is 20x those figures.

Has ChatGPT stopped accepting new subscribers?

Only for the top tier. On 10 September 2026 OpenAI paused new sign-ups for the $200 ChatGPT Pro plan worldwide because of demand for GPT-6 Astra, with no reopening date. Free, Go, Plus, Business, Enterprise and the API remain open, and existing Pro subscribers keep their access.

How do I avoid hitting Claude or Codex limits?

Start the five-hour window when the real work starts, keep threads short, use the cheapest capable model, attach only the files you need, batch heavy jobs into one session, and split work across two providers. If your work arrives in bursts, use a tool that meters by a monthly pool rather than by the clock.

Is there an AI agent without usage limits?

No serious tool is unlimited, but the shape of the limit differs. Jobbit's paid plans have no hourly, daily or weekly windows: a monthly credit pool (from 6,000 credits at £15 a month) that you can spend whenever the work needs doing, top up with credit packs, and use with Claude, GPT and Gemini models from one chat. Start free at jobbit.uk.