Guide
How LLM gateway pricing works
Short answer
LLM gateways charge in four unrelated ways: a percentage markup on tokens, a cut when you add funds to a prepaid balance, a flat fee per developer per month, or nothing at all because you run it yourself. Which is cheapest depends almost entirely on your monthly token spend, and the answer flips as you grow.
The four mechanisms
These get discussed as if they were variations on one thing. They are not. A percentage fee grows with usage and is invisible at small scale; a seat fee is fixed and is nearly all of your cost at small scale. Two products with identical headline claims can differ by an order of magnitude at your volume depending on which of these they use.
- Markup on tokensMarkup: A percentage the gateway adds on top of what the model actually costs. Some charge none at all and make money elsewhere.
- A percentage added to what the model actually cost. Charged on consumption, so it scales exactly with usage and disappears when you are idle. The most predictable of the four, and the easiest to model.
- A cut when you add fundsCredit or top-up fee: A cut taken when you add money to a prepaid balance, typically around 5%. Easy to miss because it is not a markup on tokens — but you pay it on every dollar you load.
- Taken when you top up a prepaid balance rather than when you spend it, typically around five per cent. Easy to miss because it never appears on a per-token price list, and you pay it on money you have not spent yet.
- Per developer, per month
- A flat seat charge. Independent of how many tokens you push through, which makes it punishing on a small workload with a large team and excellent on a large workload with a small one. Grows with hiring, not with traffic.
- No fee, you run itSelf-hosted: You run the software on your own infrastructure. No third party sees your traffic, and there is no vendor fee — but you own the uptime, the patching, and the upgrades.
- The permissively licensed gateways charge nothing and take no cut. You pay for infrastructure, and — the line that appears on no pricing page — for the engineer time to operate a service that is now in your request path.
A fifth thing is not a mechanism but changes the arithmetic: BYOKBYOK — bring your own key: You keep your own accounts and contracts with OpenAI, Anthropic and the rest, and the gateway routes through your keys. You keep your negotiated rates and any committed-spend discounts; the gateway charges you for the plumbing, not the tokens.. You keep your own provider contracts and the gateway routes through your keys, so you keep any negotiated rate and the gateway charges only for the plumbing. It does not lower the token price; it stops the gateway from being the one who sets it.
What each one actually costs
Four workloads, one representative of each mechanism, gateway overhead only — the token bill is excluded because it is roughly the same wherever you route it, and including it buries the difference. These figures are computed by the same model that powers the cost estimator, from the fees in the catalogue, at the moment this page was built.
| Workload | Monthly tokens | Requesty Markup on tokens | OpenRouter Cut when you add funds | Vercel AI Gateway Per developer, per month | LiteLLM No fee, you run it |
|---|---|---|---|---|---|
| Prototype 10k requests · 1 seat | $6.75 | $0.34 | $0.80 | $20 | $10 assumed |
| Early product 250k requests · 3 seats | $150 | $7.50 | $8.25 | $60 | $40 assumed |
| Scaling product 3M requests · 8 seats | $1,800 | $90 | $99 | $160 | $250 assumed |
| Heavy production 25M requests · 25 seats | $14,250 | $713 | $784 | $500 | $1,200 assumed |
The self-hosted column is the one to read carefully. Its figure is an infrastructure assumption this site makes on your behalf, not a fee anyone charges, and it excludes the engineer time entirely. Every other column is a published number.
The counter-intuitive part
Percentage fees get most of the criticism and deserve the least of it, at least early on. At the two smaller workloads above, the percentage-based options cost single-digit dollars a month while the flat costs — a seat charge, a self-hosting bill — are the dominant expense by a wide margin. A five per cent cut of a small number is a small number.
The inversion happens later, and where it happens is arithmetic rather than opinion. A flat seat fee beats a percentage cut once monthly token spend per developer passes roughly $364 — that is Vercel AI Gateway's $20/user/mo divided by OpenRouter's 5.5% cut on the money you load. Below that line the percentage wins; above it the seat does. Both figures come from the catalogue, so if either vendor changes its pricing this sentence changes with it.
Which means the fee mechanism is rarely the thing worth optimising until you are genuinely at volume. Before that point, the difference between the cheapest and most expensive option here is smaller than the cost of one engineer-day spent choosing, and you should pick on the criteria in the shortlist guide instead — deployment model, compliance, model coverage — and let the fee follow.
Which mechanism suits you
Prefer a percentage if
- Your traffic is spiky, seasonal, or still unpredictable.
- You have more developers than you have token volume.
- You want cost that falls to nearly nothing in a quiet month.
- You are still prototyping and want no fixed commitment at all.
Prefer a flat seat fee if
- Token spend per developer is comfortably above the crossover above.
- Finance wants a predictable line item rather than a variable one.
- The team is small and the workload is large.
Prefer self-hosting if
- You are choosing it for control, licence or data residency — not for the fee.
- You already run a platform team who can own another service.
- A third party in the request path is unacceptable regardless of cost.
Insist on BYOK if
- You have negotiated rates or committed spend worth protecting.
- You want the option to leave without renegotiating with model providers.
- Your security model requires the provider contract to be in your name.
What the headline number leaves out
Five charges that sit outside the four mechanisms and are routinely missed. A per-request surcharge for zero data retention, which turns a compliance requirement into a usage fee. A log-retention tier, where the useful retention window is on a higher plan than the one you priced. A minimum commitment or annual floor. Overage pricing that differs from the rate you were quoted. And guardrail or moderation features billed separately from the requests they inspect.
Each of those is a field in the catalogue, so a product that publishes one shows it on its own page with a source. 7 products in the catalogue charge no markup, no top-up cut and no seat fee under an open licence — which makes them free of fees, not free of cost. The estimator breaks any product's total into named lines and marks which ones it had to assume.
Common questions
Is a 5% credit top-up fee worse than a 5% markup on tokens?
They are close to equivalent in effect and differ in what they apply to. A markup is charged on the tokens you actually consume. A top-up fee is charged on every dollar you load onto a prepaid balance, whether or not you spend it, so unused balance has already been taxed. At the same percentage the top-up fee is very slightly worse, and both are far less money in absolute terms than most teams assume.
Is a zero-markup gateway actually free?
Zero markup means no percentage on tokens. It does not mean no cost. Look for a per-seat charge, a cut when you add funds, a per-request surcharge for zero data retention, a log-retention tier, and a minimum commitment. The estimator shows each of those as a separate line rather than one total, because which line you are paying determines how the cost grows.
Does BYOK save money?
It changes who bills you for tokens rather than reducing the token price. With BYOK you keep your own provider contracts, so you keep any negotiated rate or committed-spend discount, and the gateway charges you for the plumbing only. That is worth real money if you have a discount to protect. If you are paying list price anyway, BYOK mainly buys you the ability to leave without re-signing anything.
At what point should I self-host to save on fees?
Later than the fee percentage suggests, because the fee is not what you are comparing against. A self-hosted gateway costs infrastructure plus the engineer time to deploy it, monitor it, upgrade it and be paged for it, and that last part is the expensive one and appears in no pricing page. The honest answer is that self-hosting is usually chosen for control, data residency or licence reasons, and the fee saving is a bonus rather than the trigger.
Why do the fees in this guide differ from the vendor pricing page?
They should not. Every figure here is read from this site’s catalogue, and each catalogue value carries the vendor page it came from and the date it was checked. If you find a difference, the vendor page wins and we want to know: the corrections form goes straight to the person who maintains the row.
Do these examples include the cost of the tokens themselves?
No, deliberately. The table shows gateway overhead only, because the model tokens cost roughly the same whichever gateway forwards the request, and mixing the two makes a large identical number hide a small differing one. The estimator can show either view.