Why your AI tool says wait.

AI usage limits are a budget of computing work that refills on a timer. Claude, ChatGPT’s Codex and Google’s Gemini refill it every 5 hours and can also stop you at a weekly ceiling, while Cursor and GitHub Copilot hand out a monthly allowance. When the budget runs out, you wait, switch to a lighter model or pay for more.

So when your tool tells you to come back in a few hours, you have used up the 5-hour window. If the reset is days away, you have usually hit the weekly cap. Below are each vendor’s own rules as of September 2026, why agents drain limits so fast, and how to make an allowance last.

01What is an AI usage limit?

An AI usage limit caps how much work a model does for you over a period of time, and it is measured in compute, not in messages. Anthropic says its plans have “no fixed message count,” and Google calls Gemini’s limits “compute-based,” weighing your prompt’s complexity, the features you use and the length of your chat.

In practice, a short question barely dents your allowance. A long chat with attached files, web searches and the biggest model uses far more of it.

Tokens sit underneath all three: chunks of text about three quarters of a word long. Our guide to tokens and context windows shows how they add up. This guide is about the first kind.

02How does the 5-hour limit work?

The 5-hour limit is a short allowance that refills on its own every five hours, so one heavy morning cannot use up your whole week. Anthropic calls it a rolling five-hour session, and Settings > Usage shows how much of the current session you have used and when it resets.

Gemini’s limit “refreshes every 5 hours until you reach your weekly limit,” and OpenAI states Codex allowances per five-hour period. OpenAI is also the only vendor that puts numbers on the window.

Checked on September 28, 2026. All rows come from OpenAI’s Codex pricing page: estimated local Codex messages per 5 hours.

The ranges span roughly tenfold because the work varies. OpenAI says model choice, context, reasoning, tool use, retrieval and caching all change what a message costs, and that cloud chats may use more of the allowance than local messages.

03How do weekly limits work in Claude Code and Codex?

A weekly limit is a second, bigger ceiling that stops sustained heavy use even while the 5-hour windows keep refilling.

At Anthropic, Pro and Max plans have a weekly limit that applies across all models. It resets at a fixed day and time assigned to your account, which does not move when you start working. Claude on the web, desktop and mobile and Claude Code all draw from the same pool, so a long chat leaves less for coding.

Max plans may spend up to half of the weekly limit on Anthropic’s Fable models, while Pro users pay for Fable with usage credits. Anthropic also reserves the right to cap usage “in other ways, such as weekly and monthly caps,” to manage capacity.

OpenAI says only that weekly limits “may also apply” to Codex, and that local messages, cloud chats and ChatGPT Work share one allowance. Google applies a weekly limit to both the Gemini app and its Antigravity coding agent.

04AI usage limits by tool

Most major tools meter work in one of two ways: a 5-hour window with a weekly cap, or a monthly allowance.

Checked on September 28, 2026. Each tool name links to the page we read. The last row is Cyborb, a desktop AI agent from Orbioom and our product, which meters usage over a 5-hour window and a weekly rolling window that both reset on their own.

Plans on the same ladder differ by multiples. Claude Max gives 5x or 20x Pro’s per-session usage, ChatGPT Pro 5x or 20x Plus, and Google AI Ultra 5x or 20x AI Pro. Our AI coding plans comparison lines up what each step costs.

05Why did I hit my limit so fast?

Because an AI agent re-sends its whole working history on every step, so each turn costs more than the one before. Anthropic’s Claude Code guide lists what every turn sends: the conversation so far, your project context (instructions and every file the agent has read) and your new prompt. The conversation “grows the fastest.”

To see how fast that adds up, we modeled a 30-turn coding session in Python. A 6,000-token project context rides along on every turn, and each turn adds 4,000 tokens of file reads, diffs and replies. Real sessions vary, but the shape does not:

Part 1: input tokens processed in a 30-turn session
  one session, never cleared :  2,040,000
  cleared every 10 turns     :    840,000
  saving                     : 59%
  turn 30 alone sends        :    126,000 tokens; turn 1 sends 10,000

Caching softens the curve, since Anthropic says cached content such as project files counts less against your limits. The history still grows with every turn, though. Clearing has a cost too: the agent may need to reread a few files, so real savings are smaller than the model’s.

Four other things drain an allowance quickly:

• Bigger models. Anthropic says Opus “costs several times more per turn than Sonnet.” On OpenAI’s Codex rate card, the same message costs 5 times as much on GPT-6 Astra as on GPT-6 Sol, and 20 times as much on Sol as on GPT-6 Luna.

• Speed and images. Codex fast mode uses 2.5 times the standard credit rate, and image generation uses included limits 3 to 5 times faster on average.

• Tools and connectors. OpenAI says every MCP server adds context to your messages, and Anthropic calls tools and connectors “token-intensive.”

• Reasoning effort. Higher effort settings use more tokens, so Anthropic suggests a lower level for routine tasks.

Here is the model comparison, using OpenAI’s published credits per million tokens:

Part 2: one message with 10,000 new input, 50,000 cached input, 3,000 output tokens
  GPT-6 Astra     7.500 credits
  GPT-6 Sol       1.500 credits
  GPT-6 Luna      0.075 credits
  Astra / Sol  = 5x
  Sol / Luna   = 20x
  GPT-6 Sol in fast mode (2.5x the Standard rate) = 3.750 credits

06How can I make my usage limit last longer?

Keep sessions short, use the smallest model that does the job, and stop sending context the model does not need. Every tip below comes from Anthropic’s or OpenAI’s own guidance.

Before you change anything, list each file you plan to edit and what you will change in it, one line per file. Then wait for my go-ahead before writing any code.

Anthropic also suggests timing intensive work around your 5-hour windows, so a big job starts with a fresh allowance.

07What happens when you hit your limit?

You have five choices: wait for the reset, switch to a lighter model, pay per use, use an API key, or upgrade.

• Wait. Claude, Codex and Copilot show when your limit resets. OpenAI says its support team does not reset ChatGPT or Codex limits. Anthropic occasionally gives eligible plans a free “limit reset” that refills a session or weekly limit on demand.

• Switch models. Gemini subscribers can continue on Flash-Lite, Copilot suggests lighter models, and on Claude Max other models keep working after Fable’s share runs out.

• Pay per use. Claude usage credits bill at standard API rates, with prepaid bundles 10% to 30% off. OpenAI’s credits last 12 months, Google sells AI credits for Antigravity and Flow, and Cursor bills on-demand use at API rates. Copilot credits cost $0.01 each and Kiro’s $0.04.

• Use an API key. Codex and Claude Code can run on an API key billed per token, with no plan cap, only your bill.

• Upgrade. Each step up buys a multiple of the allowance.

08Is it cheaper to pay for extra usage or to upgrade?

As a rule of thumb, if you run out once or twice a month, paying for the overflow tends to cost less than moving from a $20 plan to a $100 one. If you run out most days, upgrade.

Cursor’s docs offer a yardstick: daily Agent users typically spend $60 to $100 a month in total usage, and power users running several agents often spend $200 or more. For more ways to trim the bill, see our guide to cutting AI costs, and for per-token rates, our LLM API pricing comparison.

09FAQ

Read next: what $20, $100 and $200 AI coding plans buy, what every AI subscription costs, and which AI model to use for each job.