Blog · 2026-07-16 · Vynaris Team
"No markup on inference" audited: what six AI gateways actually charge (OpenRouter, Portkey, Helicone, LiteLLM, Not Diamond, Vynaris)
Six AI gateways, six fee shapes, one table: worked monthly bills at $100/$1k/$10k spend, the BYOK break-even, and the two brackets where Vynaris loses. Prices verified 2026-07-16.
Every AI gateway says some version of "no markup on inference." OpenRouter says it in writing. Most of the others imply it. All of them are telling the truth, and none of it tells you what the gateway costs, because the fee almost never lives on the tokens. It lives in a monthly subscription, a payment-processing cut, a per-token routing charge, or a percentage of spend. We normalized six gateways into one table of worked monthly bills at $100, $1,000, and $10,000 of underlying spend. Prices verified 2026-07-16 from each provider's own page.
Two of those brackets are places our own gateway loses. We printed them anyway.
TL;DR
- At $100/month of inference, a flat-fee gateway like Helicone Pro ($79) is an 79% surcharge; at $10,000/month it is 0.79%. Same plan, 100x difference in effective rate.
- Percentage gateways are flat as a rate and brutal at scale: OpenRouter's 5.5% credit fee is $5.50 on $100 and $550 on $10,000.
- Vynaris self-serve (3% on the first $500/month, 1% after) sits in the middle and gets beaten on pure fee by Portkey above ~$3,900/month and by Helicone Pro above ~$6,900/month. If your bill is large and steady, a flat plan wins on fee alone.
The one table, normalized
Every gateway's fee re-expressed as dollars added on top of the same underlying inference gateway spend, plus that fee as a percentage. Not Diamond charges per token routed, so we convert at a stated blended price of $3 per million tokens and give the formula below. LiteLLM's open-source proxy has a $0 software fee; you pay for infrastructure and engineering hours instead.
Gateway Fee shape $100/mo $1,000/mo $10,000/mo
-------------------- ------------------------ -------------- -------------- ---------------
OpenRouter (credits) 5.5% Stripe, $0.80 min $5.50 (5.50%) $55.00 (5.50%) $550.00 (5.50%)
OpenRouter (BYOK) 5% after 1M free req/mo $5.00 (5.00%) $50.00 (5.00%) $500.00 (5.00%)
Portkey Production $49 flat + $9/100k req $49.00 (49.0%) $49.00 (4.90%) $49.00 (0.49%)
Helicone Pro $79 flat + usage overage $79.00 (79.0%) $79.00 (7.90%) $79.00 (0.79%)
LiteLLM OSS $0 software (self-host) $0.00 (0%) $0.00 (0%) $0.00 (0%)
Not Diamond $0.05 / MTok routed $1.67 (1.67%) $16.67 (1.67%) $166.67 (1.67%)
Vynaris self-serve 3% first $500, 1% after $3.00 (3.00%) $20.00 (2.00%) $110.00 (1.10%)Prices verified 2026-07-16. Sources: OpenRouter FAQ, Portkey pricing, Helicone pricing, LiteLLM, Not Diamond pricing, Vynaris pricing.
Three fee shapes fall out of this, and the shape decides who wins your bracket.
Shape 1: percentage of spend (OpenRouter, Vynaris, Not Diamond)
OpenRouter charges a 5.5% fee when you buy credits with a card (5% for crypto, $0.80 minimum), then passes provider prices through. Its docs are explicit: "there is no markup on inference pricing." True, and beside the point. The 5.5% is a real cost of putting money in, and it does not shrink as you scale. On $10,000/month of inference the fee is $550.
OpenRouter's BYOK path lowers this to 5% of the model list cost, and the first 1 million requests each month are free. If your request count is modest, BYOK can be close to $0; if it is high, you are back near 5%.
Vynaris self-serve is a tiered percentage: 3% on the first $500 of usage each calendar month, 1% on everything after. That is $3 on $100 (all in the 3% band) and $110 on $10,000 (the arithmetic: 3% of $500 is $15, plus 1% of the remaining $9,500 is $95). The rate falls as you spend more but never hits zero.
Not Diamond prices differently again: $0.05 per million tokens routed, independent of which model runs. As a percentage of spend, that depends entirely on the price of the tokens you route:
Blended model price Not Diamond fee as % of spend
------------------- -----------------------------
$0.30 / MTok 16.67%
$1.00 / MTok 5.00%
$3.00 / MTok 1.67%
$5.00 / MTok 1.00%
$10.00 / MTok 0.50%Route cheap DeepSeek-class tokens and Not Diamond's flat per-token fee becomes a double-digit percentage; route frontier model tokens and it rounds to nothing. The cost-per-token of your workload sets the rate.
Shape 2: flat subscription (Portkey, Helicone)
Portkey's Production plan is $49/month and includes 100,000 logged requests, with $9 per additional 100,000. Helicone Pro is $79/month; Team is $799/month. Neither marks up tokens. The fee is a fixed monthly number, which is the whole story: divided by a small bill it is enormous, divided by a large one it disappears.
At $100/month of inference, Portkey's $49 is a 49% surcharge and Helicone Pro's $79 is 79%. At $10,000/month those same fees are 0.49% and 0.79%. This is the sharpest result in the table. A flat plan is the most expensive option for a hobby project and the cheapest option for a serious one.
One caveat the percentages hide: these plans meter requests, not dollars. A workload that is cheap per call but high in volume can blow through the request allowance and add overages ($9 per 100k on Portkey) even while the token bill stays small.
Shape 3: free software, real cost elsewhere (LiteLLM)
LiteLLM's open-source proxy is free to self-host. Its software fee is genuinely $0 at every spend level, which is why it is the floor on the chart. The cost moved, it did not vanish: you run the process, patch it, add model definitions, wire up fallbacks, and carry the pager. Enterprise LiteLLM is custom-priced (contact sales, 30-day trial). Treat the $0 as a software line item, not a total, and compare it against the engineering hours a managed gateway removes. Our honest gateway total-cost comparison works that tradeoff in more depth.
Where Vynaris loses
The angle of this post required us to find the brackets where our own numbers are worse, so here they are, from the same table.
Vynaris self-serve is a percentage, so it grows with your bill while a flat plan does not. Solve for the crossover:
Flat plan Beats Vynaris self-serve above
---------------- ------------------------------
Portkey $49 ~$3,900/month of inference
Helicone Pro $79 ~$6,900/month of inferenceAbove roughly $3,900/month, Portkey's flat $49 is a smaller number than our 1% marginal rate on the excess (assuming you stay inside its request allowance). Above roughly $6,900/month, Helicone Pro's $79 does the same. If your spend is large, steady, and you do not need routing, a flat-fee observability plan is cheaper on fee than we are, full stop.
Vynaris also has an Enterprise plan at $5,000/year flat ($416.67/month, no percentage), which only beats self-serve above about $40,700/month of inference. Below that, self-serve is the cheaper Vynaris path, and we will say so.
When each shape is the right call
- Under ~$500/month, unpredictable: a percentage gateway. Flat plans are a tax when the bill is small; $79 to observe $100 of usage is not a deal. Run your own numbers before you commit to a subscription.
- $1,000 to $5,000/month, want routing: a routing gateway earns its fee by changing the spend itself, not by shaving the fee. A model routing layer that moves 30% of your calls to a right-sized cheaper model saves more than any 1 to 5% fee difference costs. That saving is not in this table, on purpose.
- Steady, high, no routing needed: a flat plan (Portkey, Helicone) or self-hosted LiteLLM. At $10,000/month, $49 flat beats every percentage option here.
- Compliance-bound, willing to self-host: LiteLLM OSS, budgeting real infrastructure and maintenance hours.
The number this table deliberately hides
A fee audit holds the underlying spend constant and compares only the surcharge. That makes it fair across gateways, and it makes it incomplete for the two gateways here that route (Not Diamond and Vynaris). Their pitch is not a lower fee; it is a lower bill, achieved by sending each request to the cheapest model that can do the job. A 30% cut to a $10,000 bill is $3,000. The largest fee gap in this entire table, OpenRouter's 5.5% versus LiteLLM's 0%, is $550 at that spend. Routing savings, when they are real for your workload, dwarf fee differences by an order of magnitude.
We will not quote a routing-savings percentage here, because the honest number depends on your traffic and we do not publish figures we cannot show you deriving. Measure it on your own mix. That is the point of the calculator and of publishing this table with our losing brackets intact.
FAQ
Does any gateway actually mark up tokens?
None of the six here mark up inference on their standard plans. OpenRouter states it plainly; Portkey, Helicone, and LiteLLM pass provider prices through; Not Diamond and Vynaris charge a routing or convenience fee separate from the token price. "No markup on inference" is true across the board and is not where the cost is.
Which gateway is cheapest at $100/month?
On pure fee, LiteLLM OSS ($0 software), then Not Diamond (~$1.67 at a $3/MTok blend), then Vynaris self-serve ($3), then OpenRouter ($5 to $5.50). The flat plans (Portkey $49, Helicone $79) are the most expensive at this spend.
Which is cheapest at $10,000/month?
LiteLLM OSS if you value the operational work at zero, otherwise Portkey ($49) and Helicone Pro ($79) win on fee because their flat cost is a rounding error at that scale. Percentage gateways cost $110 (Vynaris) to $550 (OpenRouter).
Is a flat-fee gateway always cheaper at scale?
On fee, yes, once you clear the crossover (~$3,900/month versus Vynaris for Portkey). On total bill, not necessarily, since flat plans do not route your spend down. Also watch request-based overages: a high-volume, low-cost-per-call workload can push a flat plan past its request allowance.
How is Not Diamond's per-token fee different from a percentage?
A percentage fee scales with dollars; Not Diamond's $0.05/MTok scales with tokens. On expensive models the two look similar; on cheap models the per-token fee becomes a large percentage of a small bill. See the sensitivity table above.
Why publish the brackets where Vynaris loses?
Because a fee audit that only flattered us would be worthless to you and to us. Above ~$3,900/month with no need for routing, a flat plan is cheaper on fee, and pretending otherwise would cost us the one thing this blog is for. For the underlying model prices behind all of this, see our LLM API pricing table for July 2026.
Vynaris is an OpenAI-compatible inference gateway that routes each request to the cheapest right-sized model and shows its receipts. One base URL swap. Get an API key at [vynaris.com](https://vynaris.com), and run your own workload through the [calculator](https://vynaris.com/calculator) before you switch anything.