If you keep running into usage limits on ChatGPT, Claude or Gemini, a bigger plan usually isn’t the fix. When I was starting out, I’d try all sorts of things until I struck AI gold. Once I started hitting my limits, I took a second look at how I was asking, and learned that knowing what I wanted mattered more than how I worded it. When I front-loaded context and examples, my projects hit ‘best and final’ much faster.
What a usage limit actually measures
Every plan, free or paid, is metered in tokens: the words you send, the words the model writes back, and the “thinking” it does before it answers. A five-hour session limit or a weekly cap is just a token budget with a friendlier name. Google is making that visible: in testing, Gemini’s larger context options use about 1.3 and 1.8 times the quota per turn, according to TestingCatalog.
How long a model thinks depends partly on how clear the job is and how hard you’ve told it to try. Simon Willison ran a test this month where Sonnet 5.5 at its highest effort setting thought for 128,000 tokens, about $1.28 worth, and then failed to produce the drawing he asked for. One notch lower, the same request cost under six cents and finished in 41 seconds. On Anthropic’s new small model, his same prompt cost 0.09 cents at low effort and 3.38 cents at max, about 36 times as much.
More effort isn’t the same as a better answer
Latent Space’s recap of Claude Opus 5.5 found the top effort setting cost about 2.8 times more than the medium setting and scored slightly lower, because the extra effort turned into changes nobody asked for. Vals AI tested teams of agents on the same coding benchmark: the teams cost 1.8 to 5.1 times as much, and only one setup improved significantly.
When the model doesn’t know what “done” looks like, it keeps going. It adds sections, rewrites things you liked, or hands you three options when you needed one. Each of those costs tokens, and then you spend more asking it to undo them.
Knowing what you want matters more than how you phrase it
Willison also built a feature for his blog earlier this month by talking to a coding tool while he cooked dinner. His transcript is full of “um” and “actually, no,” and it still worked, because, in his words, he “had a pretty good idea of what I wanted to build.” The prompt wasn’t polished. The picture of the finished product was clear.
The habits that burn through limits are mostly about that picture being missing:
- Asking for a piece instead of the whole job. Several small prompts that each restate the context cost more than one prompt that explains it once.
- No description of the finished product. “Write something about our new hours” leaves the model to decide length, format, tone and audience.
- Old prompt tricks. According to the Opus 5.5 recap, Anthropic’s own guidance now says to drop phrases like “think carefully,” because current models already reason and the push just makes them think longer.
- Every setting on maximum. The top effort modes are for hard problems. On everyday tasks they mostly cost more.
Ethan Mollick’s description of a good brief covers what you’re trying to accomplish and why, the limits of the job, what done looks like, what you want handed back, and what to check before finishing. He puts it simply: “what’s scarce is knowing what to ask for.” Kevin Indig and Amanda Johnson at Growth Memo suggest something even shorter: write down what winning looks like in one sentence before you start.
When I built postcards and flyers in Canva, I struggled to find the right amount of context. I narrowed my explanation to simple words, examples and uploaded sketches, and got there faster.
Your time is the bigger budget
The Growth Memo piece makes a point I’d put ahead of the token math: “Tokens are the small cost. Your team’s hours is the big one, and hours show up nowhere.” They cite a WalkMe survey in which 60% of employees said it often takes longer to figure out how to use AI than to do the task by hand. A vague ask costs you twice: once in the limit, and again in the time spent fixing what comes back.
Companies are putting budgets on their agents
Latent Space reported in September that Databricks rolled a coding agent out to about 3,500 engineers. It helped on complex work but not much on routine tasks, and coding spend went up about 60%, so they gave the agent its own budget. At a conference in New York earlier this month, Anthropic engineers said their agents don’t know their own budget while they run; developers have to set one. Bloomberg’s architecture lead listed budgets and kill switches among the controls they put around agents.
For most of us the budget is already built in. It’s the usage limit. The question is whether we spend it on the work or on the back-and-forth.
Not every limit is your fault
To be fair, sometimes it’s the vendor. Meta’s new model had a launch-day bug that drained users’ quotas far faster than normal, fixed the same day with a reset. If your limits suddenly feel tighter with no change in how you work, check the vendor’s status page before rewriting your prompts.
What I’d do
- Before you type, decide what the finished thing is: who reads it, how long it is, what format it’s in.
- Put the whole job in one prompt, with the context, instead of feeding it in pieces.
- Say what “done” means, and ask for one version unless you really want options.
- Leave the effort setting at the default, and save the top setting for problems that need it.
- If a conversation has wandered, start a new one with a better brief instead of correcting it ten times.
That’s it for now.
Sources: Simon Willison, Latent Space (AINews), Ethan Mollick, Growth Memo, Vals AI, TestingCatalog. Figures are as each source reported them, as of October 2026.
Leave a Reply