The cleanest demonstration of aggregator economics in the LLM market this year: Gemini 3.7 Flash launched August 13 at a promotional 0.75 / 3.75 per million tokens, and the same model, delivered as a batch call through a router, settles 87.5 percent below the promo price - effectively two decimals, not one. The vendor's own published promotional price was not the floor. The vendor's batch tier, aggregated and pooled, was.
Where the spread comes from
A router undercutter a vendor promo in exactly three places, and all three are standard mechanics from wholesale markets:
- Batching is a forward tier, and the vendor prices it. The 50 percent batch discount that the majors run is a price of flexibility: you lose the wall-clock, you get a tier. When a router's aggregate supply is large enough, the batch tier's discount compounds against the promo - a pooled batch cohort clears at the bottom of the vendor's own tier table, below the promo the vendor set for the interactive tier. The 87.5-percent figure is the end-to-end spread: the promo tier's price, crossed with the aggregated batch tier's clearing rate.
- Aggregation pools volume across buyers. A single tenant's batch cohort is too small to move the tier; an aggregator's is not. The aggregator's edge is the pooling - the same math as a wholesaler's edge over a single store, and it is why the aggregator's effective rate is below the rate any individual customer can get directly. This is the margin the router takes: the spread between (what the aggregate clears at) and (what it sells to you at). The 87.5-percent is the public ceiling of that spread, before the router's own cut.
- Cross-provider arbitrage on the same value. "Summarize this article" does not care which lab's mid-tier shaped the answer, and the router routes the task to the cheapest qualified model across providers. The vendor's promo competes against the router's cross-provider pool, and a vendor that promo's against direct pricing is - structurally - pricing against an opponent who sees all the other vendors' rate cards at the same time.
The margin structure, made visible
The whole thing reduces to one diagram, and it is the broker's diagram:
- The vendor sets two prices: the list price and the promo price, both tuned to the direct relationship.
- The aggregator takes the pooled cohort into the vendor's bottom tier (batch), gets a clearing rate below the promo, sells to you at a take rate above the clearing, and keeps the difference - the spread.
- You pay less than the vendor's promotional price, and the gap between your price and the vendor's promo is the sum of (the aggregator's take) and (the part of the promo's margin the vendor can no longer keep). The vendor is effectively buying the routing share - the relationship - at the spread. It is the same as the loss-leader game in the below-cost pricing post, except the subsidy is in the spread and not in the rate.
What the Stripe move does to the structure
The August 19 acquisition of OpenRouter by Stripe, reported at 7.5 billion, puts the largest aggregator inside a payments company. The economic stakes are not the model routing. They are the billing line: the aggregator that also owns the payment rail can underwrite the spread not from its routing margin alone but from the interchange-adjacent economics of the flow - a thicker margin on the same spread, which means a floor that sits further below the vendor's promo. Two watch items:
- Whether the aggregator's take rate rises, falls, or stays put post-acquisition - it is the single number that tells you the spread's center of gravity.
- Whether other aggregators respond by buy-through, or by routing around the promo tier's - the 87.5-percent is a first observation of the game, not a settled price.
What to do
- Compute your own spread per task: (aggregator effective rate on the batched version of the job) vs. (vendor promo on the same model). If the spread is real and stable, the batch + aggregator path is the default for any non-interactive traffic - that is the understanding of per-token pricing with the tier table applied.
- Keep the direct key at the promo tier for interactive traffic and for the model calls that are latency-sensitive - the spread only exists on the flexible side.
- Track the aggregator's take rate and the batch-tier clearing rates at every provider, in the price movements tracker. The spread is a variable, and the variable is the margin.