The bill does not match my numbers.
Start with the three reset times
Daily caps reset at 00:00 Beijing time; booster packs reset separately at 02:00 the next day, two hours later; each plan runs for 30 days. Refunds are not supported, but member-to-member balance transfers are, usually handled within one business day.
Four numbers that shape the bill
Daily cap reset
Beijing time. Hitting the daily cap pauses service until it resets.
Booster reset
Two hours after the daily cap, deliberately, to cover work that runs past midnight. A booster only extends the current day.
Plan validity
Each plan runs 30 days independently and must be repurchased at expiry; several plans can run at once.
Balance transfer
Refunds are not supported, but balance can be transferred between members, usually handled within one business day of the request.
Four reasons the bill does not match, in order of frequency
① Cache writes and cache reads are priced separately. On long, repeated contexts those two lines can outweigh plain input. ② Thinking tokens count as output. With thinking enabled, the visible answer length says nothing about the billed output volume. ③ One task is not one call. An agent invoking tools produces several round trips, so what felt like a single question is several charges. ④ Reset times differ. The daily cap cuts at 00:00 and the booster at 02:00, which makes traffic around midnight easy to attribute to the wrong day.
Three reset times, do not mix them up
Daily cap: resets at 00:00 Beijing time. Booster pack: resets separately at 02:00 the following day, two hours later — deliberately, to give work that runs past midnight some slack; a booster only extends the current day's allowance and does not change the plan. Weekly Opus cap, where applicable: resets on the calendar week at Monday 00:00 and applies to the Opus family only. Plan validity: each plan runs its own 30 days.
How to trace an anomalous burn
Identify the key. If the whole team shares one, this step already fails — which is exactly why we recommend a separate sub-key per person, project or pipeline.
Break down the usage: input, output, cache write and cache read separately. An anomaly usually concentrates in one of them — cache never hitting, or thinking tokens dominating.
Align the window. Confirm the day you are looking at is cut at 00:00 Beijing time, and that the booster portion is cut at 02:00. With the wrong window, no discrepancy will ever reconcile.
Confirmed terms vs common misreadings
Confirmed
① Deduction follows the usage block in the response and reconciles line by line. ② The daily cap resets at 00:00 and boosters at 02:00 — they are not in step. ③ Plans run 30 days and several can overlap. ④ Refunds are not supported; member-to-member balance transfers are. ⑤ CNY and USD balances are separate and do not mix.
Common misreadings
① "Out of quota means the plan expired" — no; hitting the daily cap just means waiting for the reset, and the plan is still within its 30 days. ② "A booster raises the daily cap" — it only extends the current day's allowance and leaves the plan unchanged. ③ "$1 of quota is a fixed token count" — no; rates differ by model, and for Claude Sonnet 5 a dollar is roughly 330K input or 67K output tokens. ④ "I can refund what I over-bought" — refunds are not supported; balance can only be transferred between members.
Two kinds of "it stopped working", handled very differently
You hit the daily cap
The day's spending ceiling is reached and service pauses until the 00:00 reset. To resume immediately, buy a booster — it extends the current day only and resets separately at 02:00. This is throttling, not an empty wallet.
Your balance is exhausted
The money itself has run out and you need to top up. Tell the two apart before acting: the daily cap is a throttle, the balance is the reservoir. Confusing them means buying a booster you did not need, or topping up when the problem was throttling.
How to reconcile a day yourself
Pick one day, sum the usage across every request for that day with input, output, cache write and cache read kept separate, multiply each by the corresponding rate on the pricing page, and compare with the bill. If the gap concentrates in the two cache lines, it is usually a caching problem rather than a pricing one. If the gap is spread evenly, check first that your window is cut on Beijing time.
On QCode
The console shows consumption per key, which is why we recommend a separate sub-key per person and per pipeline — with one shared key the bill can tell you how much was spent but never by whom. Invoices are requested separately after a top-up, with the processing fee banded by invoice value: a flat CNY 80 below CNY 2,000 and 5% at CNY 2,000 and above. Note that we do not offer refunds: a provisioned digital good occupies account resources and starts metering immediately. The alternative is a member-to-member balance transfer.
FAQ
Why do the daily cap and the booster reset at different times?
Deliberately. The daily cap resets at 00:00 Beijing time and the booster at 02:00 the next day, two hours later, to leave slack for work that runs past midnight. When reconciling, cut each one on its own window.
Does a booster raise my daily cap?
No. A booster extends the current day's allowance only; the plan's daily cap is unchanged, and the booster resets separately at 02:00 the next day. If you find yourself buying boosters often, moving up a plan is usually better value.
How many tokens does $1 of quota buy?
It depends on the model and your input-to-output ratio. For Claude Sonnet 5, a dollar is roughly 330K input tokens or 67K output tokens; cache reads are priced lower, so a workload with a high hit rate goes considerably further.
The bill shows a model I never requested. Was I switched?
Not necessarily. Many coding agents call a smaller model themselves for auxiliary work — summarising, classifying intent, compacting context — and those calls bill too. First check that the model echoed in the response matches your request; if it does and other models still appear on the bill, they are most likely auxiliary calls your client made.
Can I get a refund?
No. What we sell is a digital good: once provisioned it occupies account resources and starts metering. The alternative is a member-to-member balance transfer — email us the amount and the recipient's address, usually handled within one business day. Withdrawals are not supported.
What happens to quota left over when a plan expires?
Each plan runs its own 30 days and must be repurchased at expiry, so size your tier against what you will realistically consume rather than against the ceiling — the unused part does not carry over. Several plans can run at once, so with uneven load it is often better to stack smaller plans than to buy one large tier.
Sources
Reset times, plan validity, booster rules and token conversion: the usage rules and FAQ on /pricing and the console guidance (checked 2026-09-02). The two invoice-fee bands and the refund and balance-transfer terms: the FAQ on /pricing. Per-model rates are governed by the pricing page and move with upstream changes.
Tell a cap apart from an empty balance first
A daily cap means waiting for the reset or buying a booster; only an empty balance calls for a top-up.
Further reading
QCode pricing guide
Plan tiers, the unit-price basis and the arithmetic people get wrong.
AI API invoicing and corporate payment
The two VAT invoice fee bands, bank transfer and enterprise procurement.
How to judge whether an AI API relay is worth using
Four self-tests: multiplier conversion, cache hits, echoed model and usage reconciliation.
The reset times, validity periods and fee basis on this page reflect the rules in force on 2026-09-02 and may change; the pricing page and console at the time govern. Per-model rates move with upstream pricing.