If you have only ever used one LLM provider, the list price on that provider's pricing page probably looks like the price. It is not. It is the retail price, and like most retail prices it has a lot of margin and structure baked into it. Aggregators make their money by attacking exactly that structure.
Why direct list prices are high
Direct pricing has costs that are not about running the model. When you buy from a vendor directly you are also paying for:
•Minimum commitments. Some providers want a prepaid balance or a committed spend before they give you their best rate. Small teams rarely hit those tiers.
•One provider, one contract. Every vendor has its own signup, its own billing, its own currency. Running three providers means three invoices and three payment methods.
•No competition inside your own bill. When you are locked to one vendor, that vendor has no reason to discount. You pay their rate or you leave.
•Rate limits that are set for their protection, not your convenience.
Aggregators exist because they can buy upstream capacity in volume and pass part of the discount down. They also let you move spend between providers, which is the real leverage: if model A gets cheaper next month, you switch your traffic without switching vendors.
The math, with round numbers
Let me use deliberately simple example numbers so the shape of it is clear. Real prices change constantly, but the arithmetic is the same.
Say a provider lists a model at $2.50 per million input tokens and $10.00 per million output tokens. A typical workload, say a support-ticket summarizer doing 10 million input tokens and 4 million output tokens a month, would cost:
Direct: (10M x $2.50) + (4M x $10.00) = $25 + $40 = $65.00 / monthAn aggregator that has negotiated volume pricing might offer the same model at roughly a quarter of the list price, $0.60 per million in and $2.40 per million out:
Aggregated: (10M x $0.60) + (4M x $2.40) = $6.00 + $9.60 = $15.60 / monthThat is about $49 saved per month on one modest workload, a ~76% reduction, before you touch anything else. Scale the token counts up by an order of magnitude and the absolute savings scale with it.
The key point is not these specific numbers. The point is that the gap between "list price" and "volume price" is real, and an aggregator's job is to sit in that gap and hand some of it back to you.
The other costs that hide in the fine print
Per-token price is the headline, but it is not the whole bill. Three things trip people up:
Cache-hit pricing. Some models discount tokens that hit a prompt cache. If a provider charges full price for cached input and another charges 10% of it, a workload with repetitive prompts can differ in cost by a lot even at the same list price. Ask before you commit.
Currency and payment friction. Buying from a vendor whose billing is in a currency or region your card does not like adds conversion fees and failed-charge headaches. A single aggregator bill in your own currency removes that.
Rate limits as a hidden cost. A cheap model you can only call 10 times a minute is not cheap; you will build retry and queueing logic to compensate, and that is engineering time. Throughput limits belong in the pricing comparison even though they are not on the pricing page.
A quick cost estimator
Estimating spend before you build is straightforward once you have the numbers. Here is a small Python script in that spirit:
def monthly_cost(input_tokens, output_tokens, in_price, out_price):
# prices are per million tokens
return (input_tokens / 1e6) * in_price + (output_tokens / 1e6) * out_price
# Direct list price
direct = monthly_cost(10_000_000, 4_000_000, 2.50, 10.00)
# Aggregator price (example values)
aggregated = monthly_cost(10_000_000, 4_000_000, 0.60, 2.40)
print(f"Direct: ${direct:.2f}")
print(f"Aggregated: ${aggregated:.2f}")
print(f"Saved: ${direct - aggregated:.2f}")Run your real token counts through it with your real prices. That is the only number that matters.
Where aggregators do not save you money
I should be straight about the limits. An aggregator is a reseller. Some models are not available through resellers at all, and for some niche models an aggregator's markup can be worse than going direct. The savings are largest on the popular general-purpose models where volume discounts exist. For an exotic model you rarely call, the difference is noise.
Also, a cheaper price is worthless if the aggregator's reliability is bad. A 30% saving that buys you a flaky endpoint costs more than 30% once you factor in retries, support tickets, and your own time. Price and reliability are one decision, not two.
TokenWorks is a straightforward version of this model: OpenAI-compatible, pay-as-you-go, with a catalog of popular models from DeepSeek, Qwen, ERNIE, Doubao, Spark, and Pangu alongside GPT-4o, Claude, and Gemini, listed well under the official per-token rates.