Codex and Claude Code use the same limit architecture: a rolling five hour window, with weekly caps on top of it. The difference is what each vendor will tell you before you buy. OpenAI publishes a message count for every model on every plan, so on ChatGPT Plus you get 10 to 100 GPT-5.6 Sol messages per five hour window, 25 to 200 on Terra and 250 to 2,000 on Luna, all multiplied by five on Pro 5x. Anthropic publishes a multiplier instead, describing Max as 5x or 20x more usage than Pro without saying what Pro is in absolute terms. Neither is dishonest. But if you need to forecast before committing a team, only one of them gives you numbers to forecast with.

If you have landed here because a limit stopped you mid task, the useful thing to know first is that it was probably not the coding that spent it. On both products the agent shares its allowance with the vendor's chat product, and on the OpenAI side that sharing surface is unusually wide. More on that below, because it is the most common cause and the easiest to fix.

What are the usage limits for Codex?

OpenAI publishes them per model, per rolling five hour window, for local messages. Local messages and cloud chats share that same window, and OpenAI notes that additional weekly limits may apply on top.

Model	Plus, per 5 hours	Pro 5x, per 5 hours	Best for
GPT-5.6 Sol	10 to 100	50 to 500	Complex reasoning, ambiguous problems, advanced coding
GPT-5.6 Terra	25 to 200	125 to 1,000	Everyday production work
GPT-5.6 Luna	250 to 2,000	1,250 to 10,000	Fast, high volume, focused coding tasks
GPT-5.5	15 to 80	75 to 400	Previous generation, still selectable
GPT-5.4	20 to 100	100 to 500	Half the credit rate of 5.5
GPT-5.4 mini	60 to 350	300 to 1,750	Cheapest model on the rate card

Read the ranges rather than the midpoints. Every single row spans a factor of ten, and that is OpenAI telling you something true: the shape of your task matters far more than the plan you are on. A one file bug fix and a refactor across forty files are both "one message", and they are nowhere near the same amount of work. Model choice, context length, reasoning depth, tool use, retrieval and caching all move the number.

Underneath the message ranges sits a credit rate card priced per million tokens, which is what makes Codex spend genuinely calculable. Sol costs 125 credits per million input tokens and 750 per million output. Terra is 50 and 300. Luna is 5 and 30. That last row is the lever most people never pull: routing a simple task to Luna instead of Sol is a twenty five fold saving on input tokens for work that often does not need the bigger model at all. OpenAI adds that GPT-5.6 usage averages 5 to 40 credits per message.

What are the usage limits for Claude Code?

Structurally the same, and published in less detail. Claude Code runs against a rolling five hour window with weekly limits on top, and that allowance is shared with Claude chat and Claude Cowork. Anthropic describes the paid tiers as a multiple of Pro: Max is 5x or 20x more usage, at $100 and $200 a month respectively. What Pro itself buys in messages is not published.

Anthropic does answer the forecasting question, just from the other end. Rather than a theoretical ceiling, its cost documentation publishes measured behaviour across enterprise deployments: around $13 per developer per active day, $150 to $250 per developer per month, with 90% of users staying under $30 per active day. That is arguably the more useful number, because it describes what the software actually costs people rather than what it could cost in the worst case.

Sit those two figures next to the plan prices and something uncomfortable falls out. The observed range of $150 to $250 a month per developer is above the $100 Max 5x plan and at or above the $200 Max 20x plan. Sustained agentic coding by a working developer is a $100 to $250 a month activity, and both vendors are telling you so in their own way.

Codex vs Claude Code limits: which one gives you more?

At equal spend, nobody can answer this honestly, and you should be suspicious of any article that does. The two ladders cost exactly the same at every rung, $20 to start, $100 for 5x and $200 for 20x, but the units are not comparable. OpenAI counts messages against published per model ranges. Anthropic counts usage against an unpublished baseline. There is no conversion factor between them, so any head to head "you get X% more" claim is invented.

What you can compare is how much you can find out before paying, and there Codex wins clearly. You can price a workload on the rate card, pick a model, and estimate. With Claude Code you run a pilot and measure. Both get you to an answer; one of them lets you get there on a spreadsheet. That transparency gap is the practical finding, and it is covered in more depth in our Codex vs Claude Code comparison.

Why do my Codex or Claude Code limits run out so fast?

Four causes, roughly in order of how often they are the real one.

Something other than coding spent it. This is the big one. Codex shares its allowance with ChatGPT Work, with ChatGPT for Excel on Plus and Pro, with tasks started through Voice, and with image generation, which OpenAI says drains included limits three to five times faster than an equivalent turn without it. Claude Code shares its allowance with Claude chat and Cowork. A morning spent in the chat product is a smaller afternoon in the terminal, and almost nobody accounts for this when picking a plan.

Long sessions cost more per message. The agent carries context, and a message sent at the end of a two hour session is carrying far more of it than the first one. This is why the published ranges are so wide, and it is why "I only sent thirty messages" is not the defence it sounds like. Clearing context between unrelated tasks is the single highest leverage habit available.

You are on the expensive model by default. Sol costs twenty five times what Luna costs per million input tokens. Plenty of routine work, renaming things, writing tests for code that already exists, mechanical refactors, does not need the frontier model. On Codex you can route this explicitly. On Claude Code you have fewer knobs.

Speed settings. OpenAI notes that fast mode consumes credits at a higher rate for supported models. If you turned it on and forgot, it is quietly costing you headroom on every request.

What happens when you hit the limit?

Neither product cuts you off mid sentence. OpenAI states that if you reach your limit during an active turn, the agent can finish that turn, subject to fair use. After that you have options rather than a wall: Plus and Pro users can buy additional credits without upgrading the plan, Business and Enterprise workspaces can top up workspace credits, and any user can fall back to signing in with an API key and paying standard API rates instead. Claude Code offers the same shape of escape, letting you add usage credits to the subscription, move up a tier, or run against the Anthropic API directly.

That last option is worth understanding properly, because it changes the risk profile. Once you are on an API key, the subscription is no longer the ceiling and there is nothing structurally stopping the invoice from growing. That is fine when you have chosen it deliberately and are watching the number. It is how teams end up with a surprise at the end of the month when they have not. Any metered tool that can bill past its own subscription is worth wiring to a real-time budget alert, so the finance conversation happens on day nine rather than day thirty.

Do Codex and Claude Code limits reset?

Yes, on both, and on two clocks. The five hour window is rolling, so it refills continuously rather than resetting on the hour, which is why waiting twenty minutes often gets you moving again after a heavy burst. The weekly cap is the one that actually bites: it resets on a fixed weekly cycle, and if you exhaust it on a Tuesday you are constrained for the rest of the week no matter how long you wait between messages. Most people who describe the limits as harsh have hit the weekly cap, not the five hour one.

How do you make your usage limits last longer?

The practical list, in order of return:

Match the model to the task. The cheapest model on the Codex rate card costs a twenty fifth of the most expensive one per million input tokens. Most days, most tasks do not need the top model.
Clear context between unrelated tasks. Context is the hidden multiplier on every message you send.
Scope tasks tighter. Two well specified requests usually cost less than one vague request plus the three corrections it needs.
Check the usage dashboard weekly, not monthly. OpenAI exposes remaining limits in the CLI with a status command and on a usage page. Claude Code shows plan usage bars in its own usage view. Knowing your pace mid week is what stops a Thursday surprise.
Keep the chat product off the coding quota where you can. If your team does heavy non coding work in the same subscription, that is a real cost centre hiding inside your engineering tooling budget.

Which plan should you actually buy?

Start at the entry rung, which is $20 a month on both, and measure for a month before moving. The observed enterprise average of $150 to $250 per developer per month tells you where sustained use lands, but it also includes people running the agent most of the working day, which most developers evaluating a tool are not doing yet. Buying Max 20x or Pro 20x on day one is buying for a usage pattern you have not established.

The one exception is a team pilot with a deadline. If you have four weeks to decide and a limit will stop the evaluation halfway through, the higher tier for one month is cheaper than an inconclusive pilot. Cancel it after.

If the reason you are reading about limits is that the meter itself is the problem, that is a different question from which meter to pick. Flat priced agents exist precisely so that the forecast is a single number, and our AI coding tool pricing comparison lays out which tools in the category meter and which do not. For the full side by side on these two specifically, including the pricing ladders and who should pick which, read Codex vs Claude Code, or the detailed Claude Code pricing guide.

Figures in this article were read from OpenAI's Codex pricing documentation and Anthropic's Claude pricing and cost documentation in August 2026. Both vendors revise plans and allowances frequently, so check the current pages before committing budget.

Keep reading

Codex vs Claude Code usage limits: what you actually get, and why they run out