What people do when they run out of Claude Code or Codex

Waiting is the most common answer, and some people welcome it. The top reply in one thread of 96, with 73 upvotes: "the 5 hour window is the only break some of us get all day". Others use the gap to plan: "I keep planning, or I take a break to do something else. Creative direction is often the bottleneck in my workflow". One Codex user: "I set alarm until next reset + watch movies, do other work".

An engineer with 25 years' experience argued against taking over by hand halfway: "Taking over from an agent half way through development is quite similar to taking over from another dev, but one that can't answer questions about the work at the moment."

What it costs: nothing. Claude's 5-hour window "resets in full five hours later" than the first message in it, and on a Claude subscription Claude Code continues the task on its own after the reset. A weekly limit can be days away, which is when people look at the other choices. What you can do now has the reset rules for both tools.

The usual moves are Claude Code to Codex, Codex to Claude Code, or out to OpenRouter, OpenCode or a local model. Some split the work by tool on purpose. One person who pays for both top plans: "What helped was moving the boring file reads and long research runs to Codex, then using Claude when I actually wanted Fable."

What it costs: a second plan, and re-explaining the task to a tool that has not seen it. A short note of where the work stands, written before you switch, saves most of that. ChatGPT Plus, which includes Codex, is $20 a month. How Codex's limits work and how Claude Code's limits work say what each window gives you.

Claude's window opens on your first message (in Anthropic's words, "The period starts when you send your first message"), so people send a tiny message a few hours before they start work, by hand or on a timer, so the reset lands in the middle of the day instead of at the end. The best-known version, with 527 upvotes: "I set up a cron job to run at 7 AM every day so my 5-hour usage window resets earlier in the day."

What it costs: a little of the window, more than people expect if the message loads a lot of context. One person found "But this simple „hello world“ eats 4-5% of my usage limit." Sending it from a fresh session, with nothing else loaded, keeps it small.

People who rarely get cut off mid-task keep the window's percentage in view and wrap up before it runs out. One developer's rule: "When you get close to 85% it is time to tell Claude to "wrap-up" and update anything it learned/needs for the future into CLAUDE.md." Another splits the work and checks in between: "If the job is divided into 10 tasks, I'll ask it to complete 1-3 then check in with me. If quota looks good, continue with 4-5."

What it costs: setting up the status line once, and the habit. How to check your Claude Code usage shows how to put both windows in the status line; in Codex, /status shows them.

Compacting a large conversation is itself one large request, and resuming a long session after a break re-reads all of it. So people near the end write a short handoff note and open a new session. The most upvoted practical reply in one thread: "Don't compact, ever. Just start a new chat if you are over 400k context for example."

What it costs: a few minutes on the note. Starting a fresh session with /clear costs nothing. Why your Claude Code limit ran out so fast explains what compacting and a break cost.

The model is one of the things Anthropic lists as deciding how much a message uses. People keep the largest model for planning and hard problems and hand the rest down: "Mostly using Opus 5.5 medium on the $100 plan and delegating as much as I can to subagents using Haiku or Sonnet for tasks that don’t need Opus computing."

On Codex the difference is spelled out. OpenAI's estimates for a Plus plan run from 5 to 45 local messages per five hours on GPT-6 Astra, and from 350 to 3,000 on GPT-6 Luna. What it costs: some quality on the hardest steps, if a step that needed the larger model gets the smaller one.

Moving up a plan is often the advice people give each other. A Codex user on Plus: "I’ll be upgrading from Plus to Pro 100 pretty soon." On a company seat, the upgrade is the company's: "I was hitting session limits several times a day, and hitting weekly limits, until my employer upgraded me".

What it costs: on Claude, the pricing page puts Pro at $20 a month billed monthly, Max at $100 for five times Pro's usage every five hours and every week, and Ultra at $200 for ten times Pro's usage every week, with no 5-hour limit for verified accounts. On Codex, OpenAI says "Pro plans currently have no five-hour limit". Before paying for more, it is worth knowing what used up the window you have; why your limit ran out so fast is the place to start.

For people whose company pays, asking for more is the most common thing they say they do, and many find it tiresome: "I don't want to talk to my manager every few days to ask him to extend my limits." And: "We get $500 claude credits per month and we exhaust it in 15 days and it is painful process to get more credits."

What it costs: time, and often a written case. On a Claude Team or Enterprise plan, /usage-credits in Claude Code requests more from your admin for you. How to ask your company for more Claude Code budget has a short request to fill in.

Some people keep two or more subscriptions with the same vendor and move to the next one when a window runs out. That may break the vendor's terms of service, so read them before you try it. This page does not explain how.

Which habit pays off depends on what is using up your window: a long session, compacting, subagents or the model. OpenReason Intel, a dashboard for Claude Code and Codex usage, keeps a record of your 5-hour and weekly windows from every machine you use, with each turn's estimated share of the window, so you can see which it is and whether a change helped.

It shows you what happened; it does not stop a session as you work, and it cannot change your vendor's limits. As you work, the status line and /status are what you watch. The tools that make your allowance last covers those and others.