When free AI coding tiers reset, and what carries over

The six windows

These are the same six tools as the shortlist that includes inference, because a tier you point your own API key at has no reset to speak of. Every cell below is what the vendor publishes today, or a statement that they publish nothing.

The four unpublished rows are the same four whose allowance cannot be converted into finished work, which is less of a coincidence than it looks. A tier that counts in sessions or in "limited" requests has nothing precise to say about the boundary either, because the quantity being bounded was never precise. Publishing "00:00:00 UTC" requires having decided what, exactly, is being counted.

The cadence is the outage

Free tiers are sold on the size of the bucket. The number that actually shapes your week is the one nobody advertises: how long you are stopped when the bucket is empty.

Put that next to the marketing and the ranking inverts. A monthly allowance is the most generous thing on the page and the most dangerous shape to plan around, because the failure is a three-week outage rather than an inconvenient evening. A daily allowance is smaller and its worst case is tomorrow. Neither is better in the abstract; they are better for different working patterns, and only one of those patterns is the one you have.

Two things called a reset

Underneath the table there are two different arrangements sharing one word, and the difference decides whether the boundary costs you anything.

A refill. You are issued an allowance, you draw it down, and at the boundary it is restored to full. Anything you did not use is gone. GitHub says this in as many words: included AI credits do not carry over between months, and unused credits are forfeited. Five of the six rows above work this way, whether or not they say so.

A cap release. You hold a balance. Something limits how fast you may add to it, and the boundary lifts that limit rather than refilling anything. Clixad works this way: a free account may earn up to 500,000 credits a day, that ceiling resets at 00:00 UTC, and the credits already in the wallet are untouched by the boundary. There is no expiry date on them.

The practical difference shows up on the quiet fortnight. Under a refill, a fortnight of light work funds nothing later, because the allowance was never yours to accumulate. Under a cap release, the same fortnight leaves a balance sitting there for the week the deadline lands. Note the symmetric cost, because it is real: our earning ceiling is itself a daily one, so an account with an empty wallet and no headroom left waits until 00:00 UTC exactly like everybody else in the table. The balance carrying over is the difference, and it is the only one.

What forfeiting is worth

It is worth pricing the thing being forfeited, because the answer reframes the anxiety. GitHub Copilot Free grants 2,000 code completions a month. Here is what a full month of them costs at our metered rates, on the completion shape an earlier post declared in its markup and this page reads back out rather than restating: 1,200 input tokens and 40 output tokens each.

So a month of completions you failed to use was worth well under a dollar of inference on the models that are sensible for the job. The forfeiture is not where the money is. What the boundary actually costs you is timing: not the value of the unused half, but the fact that a light month cannot fund a heavy one, so the tier has to be sized for your worst week and is therefore idle for most of the rest.

The 50 chat requests on the same plan are the part that bites, because chat is where agent work happens. That is about four twelve-call agent tasks a month, and it resets on the same calendar-month boundary as the completions.

The paid plans document this better

The pattern in the table is worth naming, because it runs the opposite way to what you would expect from products aimed at people with less patience for surprises.

Anthropic's Claude Code documentation states that usage "resets on a rolling five-hour window and a weekly window", and that when a developer hits it, "The message shows when the window resets". OpenAI's Codex pricing page states that "The usage limits for local messages and cloud chats share a five-hour window. Additional weekly limits may apply." Neither of those is a free tier. Both publish the shape of the window and surface the reset time at the moment it matters.

Meanwhile the free tiers, whose users have the least ability to escape a blackout by paying, are the ones leaving the clock unstated. The clearest case is Google's, where the quota page gives exact per-day and per-minute ceilings with no window attached, and a user reported in March 2026 that they had "Hit the daily quota on Friday. Error shows a reset time, but 3 days later the quota is still blocked." That issue was closed as not planned. The full reading of Google's numbers is the companion to this post.

Working through the gap

Assume the window is unknown, because for four of six it is. What is left is the set of moves that do not depend on knowing it.

Drop a model tier before you drop the tool. A large share of blocked afternoons is a frontier model doing work a small one would have finished. This only helps where the tier lets you choose, which is the catch: on a request-counted quota every model costs one request, so choosing a cheaper one saves you nothing at all. On a token meter it saves you almost everything. One twelve-call task is 3,051 credits on GPT-5 Nano and 288,450 on Claude Opus 5, and that is the same task.

Find out what the stopped task already cost. An agent that halts at step 8 of 12 has spent most of the budget and produced nothing, and that share is measurable: between 53% and 55% of a clean run, on every model in the catalogue. Knowing it tells you whether to resume or restart, which is the actual decision in front of you.

Keep a second agent installed on a different funding model. Two tiers that do not share a ceiling cannot be emptied by the same afternoon. This costs nothing until the day it is the only thing working, which is the whole argument for it.

Prefer a balance to an allowance for anything with a deadline. An allowance is sized by somebody else for an average user, and it expires. A balance you can top up is sized by you. On Clixad that top-up is a completed advertiser action in a browser tab, credited on completion, which is worth stating with its failure mode attached: being screened out of a survey part way through pays nothing while still costing the minutes, and that is the most common outcome of starting one across the whole offerwall industry. Observed on our own wall in Germany, one completed action has credited between 40,000 and 300,000 credits, with payouts varying by more than 2x at the same duration.

FAQ

If the balance itself is the question rather than the clock, where free AI coding credits come from sorts the four sources by whether the grant returns, and the five routes to coding without paying sorts the tools by who ends up paying for the tokens.