Where the meter stops: what happens when AI coding credits run out mid-task

Everybody who has used a coding agent seriously has had the same afternoon. The agent is eight calls into something that was going well, and then a modal, or a red line, or a slightly apologetic sentence about your plan. What happens in the next five seconds is a product decision that almost nobody evaluates before they need it, and it is worth more than most of the features on the comparison table.

This post is about that boundary: what five documented tools do when you reach it, what your unfinished task has already cost by then, and why the check has to happen before the model generates rather than after.

Five things a tool can do at zero

These are five distinct behaviours, each documented by the vendor doing it. Nobody advertises which one they implement, because it only matters on the worst day of the month.

Hard stop

The tool refuses further work until the window resets, you upgrade, or you buy more. Anthropic's help centre puts it in one sentence: if you hit your usage limit, you will need to wait for it to reset, upgrade your plan, or purchase usage credits. It is the most honest of the five and the most annoying, because the reset runs on the vendor's clock rather than on your deadline.

Silent degrade

The tool keeps answering, on a weaker model. GitHub documents that if you use all of your premium requests, you can still use Copilot with one of the included models for the rest of the month. Nothing stops, which is the appeal and also the risk. The same prompt now gets a different quality of answer, and the changeover was a billing event rather than an editorial one.

Overage

The allowance ends and metered pricing begins, usually behind a budget you set in advance. Anthropic describes usage credits as a way to switch to consumption-based pricing at standard API rates instead of being blocked, and GitHub prices additional premium requests at four cents each. This is the only one of the five that leaves the running task unharmed. What changes is your bill, not your session.

Billed for the wreckage

Work that produced nothing usable is charged anyway, because the tokens were really spent. Linear's documentation says it plainly: failed runs, retries, or partial completions are billable for the resources used. That is not sharp practice, it is what metering means. It is also why the moment a task dies matters far more than a pricing page can tell you.

Interrupt and restart

The tool blocks on a question about your balance and loses the job it was running. A Codex CLI issue from November 2025 describes a low-credit picker that steals the terminal's focus, and notes that even after selecting an option, the original command restarts from scratch, wasting minutes in longer jobs. It was closed as not planned. This is the expensive one, and the rest of this article is mostly about why.

Where the money actually sits in a task

Take the same task the previous post priced: find why this test fails and fix it, which the agent turns into twelve model calls. It starts at 12,000 input tokens, every step adds 3,100 tokens of tool output and previous reply, and every step generates 600 tokens back. The whole thing costs 115,380 credits on Claude Sonnet 5.

The useful question is not what it costs. It is what it has cost so far at the moment it gets killed, and the answer is not proportional to how far it got, because each call carries everything before it. Which balance you were spending decides how often you meet that moment: a later post sorts the sources of free credits by whether the grant returns tomorrow or was the only one you get.

Halfway through the steps you are only a third of the way through the money. Two thirds of the way through the steps, at step 8, you have spent 53.8% of it. That share is almost a property of the task rather than of the model: across all fourteen models we price it sits between 53% and 55%, because context growth affects every price list the same way.

Which means the worst place to be interrupted is exactly where interruptions cluster. A task that runs out of budget does so late, when the calls are big, and by then the money is spent but the work is not finished. The last four calls of this task cost almost as much as the first eight.

What the restart costs

Here is the part that makes the Codex issue expensive rather than merely irritating. A restart is not a resume. The second attempt begins with an empty conversation and re-reads the same files, re-runs the same commands and re-sends every token the first attempt already paid for.

On Claude Sonnet 5: the abandoned attempt cost 62,040 credits, the clean run costs 115,380, and doing both costs 177,420. So an interruption at step 8, followed by a successful second attempt, prices the task at 1.54 times what it should have cost. Across the whole catalogue that multiplier lands between 1.53 and 1.55, from the cheapest model to the most expensive.

Two things follow. The interruption tax is real money and it is predictable, which is unusual enough in this field to be worth designing against. And a tool that keeps your conversation on disk so you can top up and continue is not offering a convenience feature. It is the difference between paying 1.54 times and paying once, plus one resend.

Why the check has to happen before the tokens

Metering is inherently retrospective. You learn what a call cost when the model has finished generating and told you how many tokens it used. Any wallet enforced from that number alone is enforced one call too late, and somebody has to absorb the difference: the user through a negative balance, or the platform through a write-off.

Linear's documentation is refreshingly direct about which way that goes, noting that balances are not updated in exact real time, so a workspace may briefly go slightly negative. That is the honest description of after-the-fact metering, and it is what a pre-authorization step exists to avoid.

What Clixad's gateway does before each call, in this order:

1. Ask what the wallet can afford, ignoring whatever max_tokens the client asked for. Otherwise a deliberately short request would read as "out of credits" and a long one as affordable.
2. Convert the balance into an output-token allowance and send it to the provider as the ceiling for that call, so generation physically stops at what the balance covers instead of overrunning it.
3. Refuse anything that cannot fund a usable reply. The floor is 512 output tokens. Below that the call would truncate mid-sentence and still cost a full round trip, so it is not started.
4. Meter the real usage afterwards and debit it, clamped to the balance as a last resort. The clamp exists so the user's balance cannot be driven negative by a provider that overran the ceiling. When that happens, the platform eats it.

Concretely, on the task above: to start call 8, which carries 33,700 input tokens, a wallet needs 10,878 credits on Claude Sonnet 5 and 284 credits on GPT-5 Nano. Below that the call is refused with a 402 and nothing is generated or charged. A balance of 5,000 credits buys no output tokens at all at that point in the conversation on Sonnet, and it is better to be told so than to receive half a sentence.

Topping up in the middle

Clixad's answer to an empty wallet is a rewarded offerwall completion rather than a card. That makes the top-up free and slow rather than instant and paid, which is a real trade and should be described as one.

Observed on our wall in Germany, credited on completion: a 3 minute action has paid roughly 40,000–42,000 credits and a 6 minute action roughly 45,000–107,000, with the top of the observed range around 300,000. Duration barely predicts the payout, which is why those are ranges rather than a rate card.

Being screened out is the normal case, not an error. The most common outcome of starting a survey is that it ends part way through and pays nothing, because you did not match the panel being bought. A product that presents that as a failure is describing the wall inaccurately.

Against the table above, one completion at the low end covers about a third of this task on Claude Sonnet 5, or seven whole runs of it on Gemini 2.5 Flash-Lite. A free account can earn at most 500,000 credits a day across every offerwall we run, which is the ceiling on this whole arrangement and belongs in the same paragraph as the good news.

Five questions before a long task

You can answer all five for any tool from its own documentation, before you find out the expensive way.

1. Which of the five behaviours does it implement at zero? A stop, a degrade and an overage are three different products, and the second is the only one that will not tell you it happened.
2. Is the running task preserved? Interrupt-and-restart costs about half a task again. Interrupt-and-resume costs one resend.
3. Does anything cap a single call? If the answer is only a monthly budget, the enforcement is after the fact and someone is absorbing the overshoot.
4. What does the tool charge for a run that produced nothing? The honest answer is "the tokens it spent", and a vendor who says otherwise is either refunding out of goodwill or not metering.
5. How long is the wait? A five-hour window and a weekly window are different products for the same money. Read which one you are on before the afternoon it matters.

What Clixad does at zero, and what it does not fix

Clixad is a terminal coding agent funded by rewarded offerwall completions rather than a subscription. At zero it does one of the five things above: it stops, per call, before generating. You get a 402 and a balance, not a truncated answer and a bill. Credits are metered at a published rate of 100,000 credits to $1 of list value, and the price of a turn is visible before you spend it.

What it does not fix: an interrupted task is still an interrupted task. Your conversation is saved on your own disk and you can top up and continue rather than starting over, so you pay one resend instead of the 1.54 multiplier. You still pay that resend. And earning the credits back means completing an advertiser action that may screen you out, which is time rather than money but is not nothing.

The daily cap is the other honest limit. On Claude Opus 5 the most a free account can draw in a day is under two tasks of the kind priced here. That is a real ceiling, published on purpose, because an article about other people's boundaries is worthless if it hides its own.

FAQ