Most articles about AI credits ask how far they go. The more useful question is what breaks when they stop — because the answer is not the same everywhere, and on at least one tool it is considerably worse than “you cannot generate anything new today”.
Two very different kinds of empty
When a credit balance hits zero, one of two things happens.
In the first, you stop being able to make new things. Whatever you already made still works. The cost of running out is inconvenience and a top-up.
In the second, something you already shipped stops running. The cost of running out is downtime on a live service, and it arrives through your billing page rather than your monitoring.
The difference matters enormously and is almost never on the pricing page.
Lovable: the deployed app pauses
Lovable builds an app and can also host it, including a built-in backend it calls Lovable Cloud. One credit balance pays for all of it — building, chatting, hosting, the backend and AI features inside your deployed app.
Its documentation is explicit about zero: “When your workspace has no credits left, building stops, AI features in deployed apps stop working, and apps that use the built-in backend (Cloud) pause.”
So the app you launched last month goes down because this month’s allowance ran out. That is a sensible engineering decision — they are not going to run your backend for free — and Lovable documents it plainly. But it changes what a credit balance is. It is not a budget. It is a dependency of your uptime.
If you are using Lovable to prototype, this is irrelevant. If you have put anything in front of real users on Lovable Cloud, the credit balance belongs on whatever list you keep of things that can take the service down.
You pay for the attempt, not the result
There is a second clause on the same tool worth knowing before you budget. From Lovable’s terms, last updated 28 August 2026:
“Credits are consumed by each AI action based on the effort and resources used, regardless of the outcome, and are non-refundable and non-restorable even where the AI Output is erroneous, incomplete, or must be regenerated.”
Read in isolation that sounds harsh. It is also the only honest way to meter a service where the compute is spent whether or not you liked the answer. Every credit-based AI tool works this way in practice; Lovable is unusual in saying so in its terms rather than leaving you to find out.
What it means for planning is simple. A feature that takes four attempts costs four times the credits, and the three failures are not refunded. Your real cost per working feature is the advertised cost multiplied by however many goes it takes, which is not a number anyone can give you in advance.
Grok: you keep going, for a fee
Grok sits at the other end. Every tier includes a pool of usage shared across Grok and Cursor, and xAI’s own FAQ says that if you reach your included usage “you can keep going with top-ups, or wait for your allowance to refresh”. The allowance refreshes weekly rather than monthly, so a heavy day costs you days, not the rest of the month.
Grok is not hosting anything for you, so there is nothing to pause. Running out means you stop generating until you pay or until the week turns over. That is the benign version.
What Grok does not tell you is where the ceiling is. Paid tiers are described only as “4x”, “15x” and “50x” more usage than Plus, and Plus itself is never quantified. You will find out what your limit is by hitting it.
Four questions to ask before you commit
- Does anything I have already shipped stop working at zero? If the vendor hosts your app, database or backend, assume yes until the documentation says otherwise.
- Are failed or regenerated attempts charged? Almost always yes. Budget for the number of attempts, not the number of features.
- Does the allowance roll over, and when does it reset? Daily grants that expire nightly behave very differently from a monthly pool. Lovable’s daily grants expire at end of day and never roll over; Grok’s pool refreshes weekly.
- Can I cap spend before it happens? Lovable offers per-member and per-project credit limits with alerts on paid plans. Being able to set a ceiling is worth more than a generous allowance you cannot see.
The free-tier trap underneath all of this
Credit systems make free tiers hard to read. Lovable’s free plan advertises five daily build credits, which sounds like daily access. Its documentation explains that the daily grant is capped at thirty a month, and spells out the consequence: “on Free, you receive daily build credits for the first 6 days each month (6 × 5 = 30), then none until the next month starts.”
Six days, not thirty. If you were planning to evaluate the tool over a fortnight, half of that fortnight has no build credits in it.
What we checked
Lovable’s pricing page, credit documentation and terms, and xAI’s pricing page and FAQ, were read on 3 November 2026 and quoted directly. We have not tested either product, and nothing here describes how the tools perform — only what their published documents commit them to.
For what credits actually buy across the rest of the directory, see What AI credits actually buy, tool by tool. For how the same three pricing pages handle the numbers they do publish, see Three pricing pages, and the number each one leaves out.