Empower growth and innovation with the latest AI Dev insights

Client prepaid a full year of API credits at the old price, then the upstream vendor repriced mid-year — for 2026, should the bill follow the top-up date or the usage date?

Nov 16, 2026 Read: 8

Conclusion first: For AI API aggregation, gateway, or resale work in 2026, when a client has prepaid a full year of credits at the old price and the upstream vendor reprices mid-year, the steadier handling is to keep two tracks separate — calls that have already happened settle against the price list version in effect at the time of the call; how the unconsumed balance converts back is explained by the contract pricing terms at the time of top-up. Segmented settlement requires that the ledger actually stores a unit-price version layer. If it does not, the only options are a negotiated one-off discount, or carrying the difference into the next billing period as an offset.

Unpacking that sentence, it is really answering two different questions: the consumed part is "how do we record historical fact," and the unconsumed part is "how do we interpret the contract." Discuss them together and both you and the client get muddled.

1. Why "top-up date or usage date" gets argumentative so easily in resale

Token spread businesses make gross margin from the gap between upstream cost and downstream sale price. Model families such as GPT, Claude, Gemini, Qwen (Tongyi), and DeepSeek each reprice on their own cadence, and one or two price changes a year is a common situation. The client puts in money and receives credits, with the unit price sitting in between; once the unit price moves, "how many credits is the money I put in at the start of the year still worth" turns from a technical question into a contract-interpretation question.

The bigger problem is that three sets of books use misaligned coordinate systems: upstream bills by token, and input, output, cache hits, and batch processing often have different prices; the gateway records by request; the reseller sells to the client by plan or credit package. Each set does its own arithmetic, so the "how much is left" the client sees and the "how much was used" the upstream sees are not on the same dimension, and either side can produce the book that favors them.

  • The client looks at balance, the upstream looks at tokens, and the layer missing in between is the unit-price version — leaving the dispute to be settled by someone's judgment call.
  • Input price, output price, and cache price for the same model family differ by multiples, so before discussing timing conventions you must first pin down which one is being counted.
  • Multimodal requests are priced per image, per second, or per call and cannot be converted directly to text tokens; recalculating them mixed together easily fails to reconcile.

2. The four-layer metering chain: which layer does the timing convention land on?

Rather than digging back through the books after a dispute, split metering into four layers at launch — call it the four-layer metering chain: capability layer → business layer → carrier → data and risk control. The reasoning is straightforward: a price change should only touch the business layer's price list, not the capability layer's invocation method, and certainly not make the balance shown on the carrier jump on its own. Each layer has one easy-to-miss point:

  1. Capability layer (models and multimodal routing): record the model family, modality type, context tier, and whether the cache was hit; store "model alias" and "upstream real model" separately, or historical books will not reconcile after the upstream renames something.
  2. Business layer (pricing and plans): version the price list, and state the billing unit (per thousand tokens, per call, per second, per image), discount ratio, and effective timestamp; you cannot keep only a "currently in effect" row — historical versions must be retrievable.
  3. Carrier (console, mini program, app, H5): the balance, line items, and invoice basis the client sees all come from the same ledger; do no secondary conversion on the carrier, and put all conversion in the business layer.
  4. Data and risk control (usage ledger and reconciliation): each request logs request ID, unit-price version, billing unit, end reason, and time window. Without the unit-price version, any later day-splitting algorithm is only an estimate.

Layer four is the one most often skipped. It does not feel like a problem at the time — until a client brings a screenshot asking "why was this charge higher than last month," and you find the only thing you can give them is a total.

3. Delivery scene: one price change, two people spent two to three weeks rebuilding the ledger

The constraints were typical: a small team with a self-built gateway that, to save tables, stored only the charge amount and volume per request at launch, not the unit-price version. The client was a resale operator with dozens of prepaid customers, annual prepaid amounts in the range of tens of thousands to hundreds of thousands, billed monthly. After a commonly used model family repriced in 2026, one client that had prepaid at the start of the year used official page screenshots to demand the remaining balance be converted at the new price, and several others quickly asked the same question.

The response was not to agree on the spot to make up the difference, but to pull a range by request ID and reconcile: records that could surface a unit-price version settled in segments by effective point, while those that could not were only estimated at that month's weighted average price, and that portion was listed separately as a reconciliation attachment. The result was a chunk of difference between before and after recalculation that nobody could explain clearly; the client would not accept a one-off discount, and in the end a price-change clause was signed, agreeing on the old price before the effective point and the new price after. The cost, in our experience, was two people spending two to three weeks on data reconciliation and explaining the convention — clearly above the investment of building one extra price-list version table at the start.

4. Three timing conventions, compared side by side

"Which date to count from" can be implemented three ways; the difference is not sincerity but which one your ledger can support. Before choosing, confirm one thing: can historical bills pull a unit-price version per request? If yes, go segmented; if no, negotiation or a next-cycle offset are the only options.

  • Convention A — lock everything at the top-up date price: both consumed and unconsumed amounts are calculated on the price list in effect at top-up. Suits projects that launched recently with request volumes in the hundreds of thousands; reconciliation effort is in the experience range of 3–10 person-days. The trade-off: if the upstream keeps cutting prices, the longer the lock period the more gross margin is squeezed — how much varies widely by model family and discount, with a typical range from single digits to the mid-teens of a percentage point, and should be measured against your own price list.
  • Convention B — segment by time of call: old price before the effective point, new price after; the difference lands only on prepaid unconsumed balance, converted into credits or an offset credit with a clearly stated validity period. Applies to most resale scenarios; requires the price list to carry effective timestamps, and the explanation cost is relatively low.
  • Convention C — discount tracks in real time: the contract simply states "upstream real-time price × discount," so the price naturally follows the upstream and there is no timing dispute at all. Suits scenarios with stable call volume and clients who accept unit-price fluctuation. The trade-off: bill unit prices will move, and customer-service explanation volume is generally a bit higher than under a locked-price model.

The three are not mutually exclusive: a common practice is C for new clients, B for existing prepaid clients, and A only for small-amount, short-billing-cycle cases.

5. Where it applies and where it does not

Segmenting by time of call fits when the client is on prepaid, call volume is moderate, the contract states that prices may adjust with the upstream, and the gateway can trace each request back to a unit-price version. For these projects, building the price-list version and effective timestamps into the backend saves far more effort than explaining case by case afterward.

The boundary should also be explicit: if the contract locks the price for a defined term, a client asking for a recalculation based on the price cut lacks a contractual basis, and usually no make-up is owed. Conversely, if you sell "upstream price × discount," a unilateral upstream increase is also hard to push onto clients — it depends on how the price-change and notice clauses are written. One quotable boundary: whether to count from the top-up date or the usage date depends on the contract's pricing basis and the ledger's audit trail, not on client sentiment, and not on how much the upstream cut prices.

There is also a category that does not need this mechanism: an internal tool with a single client, a single model, and very low call volume (in our experience a few hundred calls a month or less). Monthly or flat-rate billing is simpler; forcing price-list versioning just adds burden.

FAQ

If the upstream cuts prices after a client prepays, does segmenting by time of call mean making up the full difference?

No. Segmentation only affects calls after the effective point and the prepaid unconsumed balance; earlier calls still settle at the price in effect then, and the difference is usually much smaller than a full recalculation.

Should the bill show the client the price-list version number and effective time?

Yes, it is worth marking. When clients can match "from when, and at which price, was I charged," most disputes end at the explanation stage — and you get a notice record for free.

When the upstream raises prices, can we apply the new price directly to clients who already prepaid?

Depends on the contract. If it says "adjusts with the upstream," you can apply it from the agreed effective point; if it locks the price, you usually wait until the term ends. Advance notice periods in practice are commonly 15–30 days, but the clause governs.

Should the gateway alert on price-list changes?

Yes. After a change, automatically diff against the previous version and notify affected clients — cheaper than explaining case by case afterward, and it leaves a notice evidence trail for reconciliation lookups.


If you are running an API aggregation or resale project in 2026, start with three small things: read through the pricing basis in your contracts, add version numbers and effective-time fields to the price list, and add request-level unit-price versions to the ledger. In API gateway and reseller backend deliveries, Xiyue Company usually places these three things before launch rather than after a dispute. The boundary also bears stating: for internal tools with very low call volume and a single client, there is no need to add this complexity.

Interested in this topic?
10-year tech team — reference proposal within 24 hours
Obtain Proposal
Are you ready?
Then reach out to us!
+86-13370032918
Discover more services, feel free to contact us anytime.
Please fill in your requirements
What services would you like us to provide for you?
Your Budget
ct.
Our WeChat
Professional technical solutions
Phone
+86-13370032918 (Manager Jin)
The phone is busy or unavailable; feel free to add me on WeChat.
E-mail
349077570@qq.com
Submitted successfully
Thank you for your trust. We will contact you soon!
Recommended projects for you