TL;DR (July 2026): If your product calls LLMs on behalf of customers, you have exactly three ways to rebill the cost: pass-through (charge what the provider charged you, 1.0x), markup (pass-through plus a margin), or flat (a fixed price that absorbs the variance). This month the decision went public: PostHog shipped an open pull request billing its Code product's LLM usage as pass-through credits at 1.0x, explicitly no markup - a break from its standard 20% AI margin - while at the other pole a $6/month "unmetered, no token tracking" API pitched on Hacker News and flat per-call gateways sell against metering itself. Each model is a different bet on trust, margin, and variance - and each one puts a different, concrete requirement on your metering pipeline. Pick the rebilling model first; the meter you need falls out of the choice.
Every AI product that resells model calls is quietly a billing company. The provider invoices you in tokens; your customer expects a price in something they understand. The translation between those two is the rebilling decision, and in 2026 it has become a visible piece of product strategy rather than a back-office detail.
The three models, and who is running each
| Model | What the customer pays | Who eats the variance | Seen this month |
|---|---|---|---|
| Pass-through (1.0x) | Exactly the provider cost | Nobody - variance is forwarded | PostHog Code: LLM costs as pass-through credits, no markup, in a public PR |
| Markup (cost-plus) | Provider cost times 1.x | Customer pays variance plus margin | PostHog's own standard 20% margin on its other AI features |
| Flat (absorb it) | Fixed per call or per month | The vendor | A $6/month unmetered LLM API on Show HN; flat per-call gateways like AkaRouter |
None of these is the correct answer. They are three different trades:
- Pass-through is a trust play. Charging 1.0x says the AI is not your profit center - you monetize elsewhere and refuse the temptation to margin an input your customer can price-check against public rate cards. In a year when metered AI billing is breaking developer trust, showing your cost basis is a differentiator. The price: zero margin for error. Every token you fail to attribute to a customer is money you personally lose.
- Markup is honest cost-plus - if you disclose it. A stated 20% covers retries, failed calls, orchestration overhead, and the metering infrastructure itself. Undisclosed markup, by contrast, is exactly what the gray market is arbitraging (more below).
- Flat is a variance insurance business. The vendor absorbs token variance and prices for the cohort average. It sells brilliantly - predictability is what buyers say they want - but it is the model that dies by its heaviest users. Flat-rate AI plans collapsed into caps across the industry for one reason, which we covered in Unlimited AI plans are dead: agents made the tail of the usage distribution unboundedly expensive.
What each model demands from your meter
The rebilling choice is also an engineering spec. Working through what each one requires:
- Pass-through needs perfect per-customer cost attribution. At 1.0x, your rebill is only as accurate as your attribution. That means capturing, per customer: input and output tokens per call at the correct per-model rate, cache reads versus cache writes (billed differently, and large), retries and failed calls (you paid for them - decide explicitly whether the customer does), and every sub-agent or tool call an agent run fans out into. This is the full attribution problem, with money directly on the line for every unattributed event.
- Markup needs everything pass-through needs, plus margin telemetry. Your realized margin is not your stated margin: provider price changes, cache behavior shifts, and retry storms all move it. A meter that reports realized margin per customer per model is how a 20% markup stays 20%.
- Flat needs cohort economics and a kill-switch. No per-call invoice, but you must know your P95 user's cost, watch for the customer whose usage makes them unprofitable, and enforce a fair-use boundary before they sink the plan - the machinery of hard caps and usage kill-switches, pointed inward at your own margin instead of outward at a provider bill.
In all three cases the meter is not optional accounting - it is the pricing model's load-bearing component. The difference is only what it must prove: cost (pass-through), margin (markup), or cohort survivability (flat).
The gray market under the list price
One more thing this month's research surfaced: a visible relay market reselling model access below official rates - one gateway advertising Chinese models at up to 50% under official pricing through a single API key, another offering 34 models paid in USDT. Whatever one thinks of the arbitrage, it is evidence that undisclosed spread gets competed away fast now. If your rebilling depends on customers not knowing what tokens cost, that is a short-lived position; public rate cards and cheap relays have made the cost basis common knowledge. The durable versions of AI rebilling are the transparent ones - 1.0x with visible cost, or cost-plus with a stated margin for a stated service.
Choosing, in practice
- AI is a feature of your product, monetized elsewhere: pass-through buys trust and neutralizes price-shopping. Budget for the attribution engineering; it is the whole cost of the model.
- AI is your product and you add real orchestration value: disclosed markup. State the margin, meter your realized margin, and let the delta between them page someone.
- Your users are light and homogeneous and predictability is the pitch: flat with a fair-use boundary - and re-examine the cohort math every time agents get more capable, because the tail grows.
The honest take
PostHog putting "no markup" in a public pull request is the interesting datum: rebilling transparency is becoming something vendors compete on, not something they hide. The uncomfortable corollary for anyone building on LLMs is that all three viable models assume you can account for every token a customer caused you to spend - cache reads, retries, sub-agents and all. The rebilling strategy is a pricing decision; the meter underneath it is just the truth. Get the truth right first and every one of the three models is available to you; get it wrong and none of them are.