# AI gateway pricing models: markup, fees, requests, subscription

URL: https://infro.io/blog/ai-gateway-pricing-models-compared
Author: INFRO Editorial (https://infro.io/blog/editorial)
Publisher: INFRO
Published: 2026-10-13
Last modified: 2026-10-13
Last reviewed: 2026-10-13 (next review due 2026-11-12)
Topic: Cost and FinOps
Format: Comparison

> AI gateways charge on one of six bases: a percentage of usage balance you buy through them (OpenRouter, Cloudflare Unified Billing, Eden AI), a markup on model cost (Requesty), a fee on your own-key usage above an allowance (OpenRouter BYOK), request, log or model volume (Kong, Portkey, Vercel add-ons), tokens routed (Not Diamond), or a flat subscription. Open-source gateways such as LiteLLM charge nothing for the software and cost you the infrastructure and engineering to run them. Whatever the gateway charges, your provider's charge for model usage is a separate figure.

Key takeaways:
- Compare gateway charges by what the fee is charged on: purchases, model cost, your own-key usage, requests, tokens, or nothing that moves with traffic.
- A percentage fee grows with model spend; a request-based price grows with traffic; a subscription changes only when you change plan.
- OpenRouter's BYOK fee above its allowance is computed on list price, not on the rate your contract gives you.
- Buying usage through a gateway's account means paying that gateway's rate for the model; bringing your own key keeps your provider's contract rate.
- Self-hosting has no licence fee and a real cost in servers, upgrades, security patches and on-call time.
- Every fee here was read on 2026-10-13 and will change; re-check the vendor's own page before deciding.

Most AI gateway comparisons line up features and leave price to a footnote. The price deserves more attention than that, because two gateways with similar features can charge on entirely different things. One takes a percentage of the money you route through it. Another charges per request. A third charges for the software and nothing else. Over a year, the base of the fee decides more than its size: a percentage grows with every model price rise and every new workload, a request price grows with traffic, and a subscription moves only when you change plan.

This article lists how a set of gateways charge, each fee quoted from the vendor's own page on 2026-10-13, says what each fee is charged on, and works through one illustrative month. It covers charges, not features or performance.

## Six bases a gateway can charge on

[Diagram: Gateway charges grouped by the quantity they move with.]

1. **A percentage of purchases.** You buy a usage balance from the gateway, which calls the provider on its own account, and the gateway takes a fee when you buy.
2. **A markup on model cost.** The same arrangement, with the percentage applied to each request's model cost.
3. **A fee on your own-key usage.** You bring your own provider key; the gateway charges a percentage of that usage, sometimes above a free allowance.
4. **Request, log or model volume.** A price per million requests, per hundred thousand logged requests, or per model proxied.
5. **Tokens routed.** A price per million tokens that pass through a routing service.
6. **A flat subscription.** A fixed monthly or annual price per plan, with allowances.

Open-source gateways add a seventh answer: nothing for the software, and everything for running it.

One distinction runs underneath all six. When you buy usage through a gateway's account, the gateway's rate for the model is what you pay, usually the provider's list price. When you bring your own key, your provider bills you directly under your own contract, at whatever rate you negotiated, and any gateway fee is a separate charge on top.

## What each gateway charges

### OpenRouter

OpenRouter sells model access through its own accounts. Its pricing page lists a platform fee of 5.5% on the pay-as-you-go plan and 8% on Business, and its FAQ says the fee "is charged when you buy credits, 5.5% on Standard and 8% on Business, never on individual requests"; Enterprise fees are set by contract[1].

Bring-your-own-key traffic has its own terms: "$25,000 of list price inference / month with no fees, 5% fee after"[1]. The BYOK guide says the fee is "5% of what the same model/provider would cost normally on OpenRouter", measured by cost rather than request count, and taken from your OpenRouter credits[2]. The base matters: the fee is computed on list price, not on the rate your provider contract gives you. By default, when your keys fail or hit a rate limit, requests fall back to OpenRouter's shared capacity, which spends credits; a per-provider setting turns that off[2].

Stripe agreed to acquire OpenRouter (announced 2026-08-19); as of 2026-10-13 neither company had published a completion notice[3]. Pricing direction after a completed acquisition is not something either company has published.

### Vercel AI Gateway

Vercel says its gateway "charges no markup and no platform fee on tokens", and that you pay the provider's list price, deducted from purchased AI Gateway credits; "you're responsible for any payment processing fees that may apply", which invoiced Enterprise billing avoids. BYOK carries no markup or fee, but needs purchased credits, and a failed BYOK request is retried with Vercel's own credentials and charged to your credits balance[4].

Some capabilities are charged per request as add-ons: a team-wide zero data retention setting at $0.10 per 1,000 requests, a team-wide provider allowlist at $0.10 per 1,000 successful requests, custom reporting at $0.075 per 1,000 tag writes and $5 per 1,000 queries, and trace drains at $0.05 per 1,000 traces plus $0.50 per GB of egress. The per-request versions of zero data retention and the provider filter carry no extra charge[4].

### Cloudflare AI Gateway

Cloudflare says the gateway's core features (analytics, caching and rate limiting) are free; logging follows Workers Logs pricing for customers who created their first gateway on or after 2026-09-24[5]. If you buy model usage through Cloudflare's Unified Billing, "a 5% fee is applied to all credits purchased", so a $100 purchase costs $105, and provider inference is passed through "with no markup"[6]. The Unified Billing page documents no fee on bring-your-own-key traffic.

### Requesty

Requesty's pay-as-you-go plan is "5% markup" on model cost: a model costing $10 per million tokens at the provider costs $10.50 through Requesty. A free plan allows 200 requests a day on free models; Enterprise is custom. Bring-your-own-key is listed as a feature with no separate fee stated[7].

### Eden AI

Eden AI says it "charges no markup on provider pricing", adding "a 5.5% platform fee applied at checkout" when you buy credits. Its self-serve plan is pay-as-you-go; its advanced plan is custom. The page does not mention bring-your-own-key[8].

### Kong AI Gateway

Kong charges for the API platform, not for model spend. Konnect Plus starts at $25 a month with one million requests included and $200 a month for each additional million; five distinct LLMs are included, and each further model is $100 a month, counted hourly and enforced monthly. Kong's paid AI plugins, including AI Rate Limiting Advanced, are an add-on on Plus and included in Enterprise, which is custom; fully self-hosted gateways are part of a separate Gateway Enterprise offering[9].

### Portkey, now Prisma AIRS AI Gateway

Portkey's pricing page now states that "Portkey is now PRISMA AIRS AI Gateway". It lists a free Developer plan with 10,000 recorded logs a month, Production at $49 a month with 100,000 recorded logs and "+$9 overages per additional 100k requests", and custom Enterprise pricing. The page mentions no fee on model spend[10].

### LiteLLM

LiteLLM's open-source gateway is free software you deploy yourself. Its Enterprise offering adds governance, SSO and SCIM, air-gapped deployment and support with response-time targets, at custom pricing with no published price; a 30-day self-hosted trial key is issued by email[11].

### Not Diamond

Not Diamond is a model router rather than a gateway. Its pay-as-you-go price is "$0.05 per million tokens routed", with custom Enterprise pricing. Its FAQ says it integrates with your existing gateway; the page does not say whose provider account the model calls run on[12].

### Self-hosting any open-source gateway

No vendor fee at all. You pay for the servers, the database and cache it needs, upgrades and security patches, and the engineers who are paged when a service in every request path misbehaves. That cost is real, and it does not appear on any pricing page.

## Side by side

| Gateway | How it charges | What the fee is charged on | Your own provider keys |
| --- | --- | --- | --- |
| OpenRouter | 5.5% (pay-as-you-go) or 8% (Business) platform fee | Credit purchases | $25,000 a month of list-price usage free, then 5% of list price |
| Vercel AI Gateway | No markup or platform fee on tokens; per-request add-ons | Add-ons: requests, writes, queries, traces | No fee; purchased credits required |
| Cloudflare AI Gateway | Core features free; 5% on Unified Billing | Credit purchases through Unified Billing | No fee documented |
| Requesty | 5% markup | Model cost | Listed; no fee stated |
| Eden AI | 5.5% platform fee | Checkout (credit purchases) | Not mentioned |
| Kong AI Gateway (Konnect Plus) | From $25 a month; $200 per extra 1M requests; $100 per extra model | Requests and models proxied | Your keys |
| Portkey / Prisma AIRS | $49 a month (Production); $9 per extra 100k requests | Recorded logs and requests | Not stated on the pricing page |
| LiteLLM | Open source free; Enterprise custom | Nothing, or a contract | Your keys |
| Not Diamond | $0.05 per million tokens | Tokens routed | Not stated |
| Self-hosted | No vendor fee | Your infrastructure and engineering | Your keys |

Table: Quoted from each vendor's own page on 2026-10-13. Fees change; re-check before deciding.

## One month, worked through

The arithmetic below applies each published price to one illustrative workload. It is not a ranking: these products do different things, and a fee buys different capabilities in each.

> **Example: An illustrative month of gateway charges**
> Assume a team whose model usage is worth $20,000 a month at the providers' list prices, across 2 million requests, 8 distinct models and 4 billion tokens. In every row, the providers' charge for that usage is a separate $20,000 (or less, under a negotiated contract on your own keys) and is not included.
> 
> - OpenRouter, buying usage through OpenRouter: 5.5% × $20,000 = $1,100 in fees.
> - OpenRouter, your own keys: $20,000 of list-price usage is under the $25,000 allowance, so $0. At $40,000 of list-price usage: 5% × ($40,000 − $25,000) = $750. If your contract rate is 80% of list, your providers bill you $32,000, and the fee is still $750, because it is computed on list price.
> - Cloudflare, buying through Unified Billing: 5% × $20,000 = $1,000. On your own keys: no fee documented.
> - Requesty pay-as-you-go: 5% × $20,000 = $1,000.
> - Eden AI: 5.5% × $20,000 = $1,100.
> - Kong Konnect Plus: $25 + $200 for the second million requests + 3 × $100 for models six to eight = $525, before any AI plugin add-on.
> - Portkey Production: $49 + 19 × $9 for 1.9 million requests beyond 100,000 = $220.
> - Vercel on your own keys with team-wide zero data retention on: 2,000 × $0.10 = $200, plus payment processing on the credits you must hold.
> - Not Diamond routing every token: 4,000 × $0.05 = $200, on top of whatever gateway it routes through.
> - INFRO Team plan: $199 a month, or $165 a month billed annually; 2 million requests are the Team allowance.

[Diagram: The gateway-side figures from the example above, on the same illustrative month.]

Two things change the picture more than the percentages. First, the base: double the model usage in the example and every percentage row doubles, while the request-priced and subscription rows do not move. Double the requests at the same spend and the opposite happens. Second, the rate you pay for the model: buying through a gateway's account means the gateway's rate, usually list; your own key means your contract.

[Diagram: Keep the provider's charge and the gateway's charges as separate lines; they answer different questions.]

## INFRO's pricing: a subscription, and nothing on model spend

INFRO is a hosted gateway over provider accounts you connect. Your providers bill you directly for model usage, at your own rates; INFRO charges a software subscription and takes 0% of model spend. Nothing INFRO sells is priced per token, and provider spend never appears on an INFRO invoice.

| Plan | Monthly | Billed annually, per month | Requests a month | Seats | Projects |
| --- | --- | --- | --- | --- | --- |
| Developer | $35 | $29 | 100,000 | 1 | 3 |
| Startup | $79 | $65 | 500,000 | 3 | 10 |
| Team | $199 | $165 | 2,000,000 | 10 | 30 |
| Business | $599 | $499 | 10,000,000 | Unlimited | Unlimited |

Table: From INFRO's plan definitions on 2026-10-13. Annual plans are paid upfront. Enterprise is by agreement.

- **No free plan.** Every plan has a 14-day trial with a card, and a trial is the chosen plan in full.
- **Overage is measured, not billed.** Past the request allowance, requests keep being served and counted; the billing screen shows it and offers the larger plan. No charge is raised for the excess.
- **Provider cost stays separate.** Each request carries an estimate of what your provider will bill you, from your own rate if you entered one, else the vendor's list rate, or unpriced where no rate is known. It is never added to the subscription. Budgets on every plan cap that estimate; entering your negotiated rates is included from the Team plan; reconciling the estimate against OpenAI's and Anthropic's own cost reports is on Business.

## Fair comparisons

A subscription is not automatically the right answer, and neither is a fee on spend.

- **Gateways that sell model access through their own accounts** remove the need to hold an account, a contract and a key with every provider. For a small team trying many models, a percentage on purchases can be a reasonable price for that convenience.
- **Self-hosting** has no licence fee and full control of the data path. It also needs people who will run, patch and page for a service every request depends on. Count that time before comparing it with any invoice.
- **Request-priced gateways** suit high-value, low-volume traffic; percentage fees suit low spend with many requests. The example's two doublings show why.

Whichever you choose, ask each vendor three things and get the answers in writing: what the fee is charged on, whether your own-key traffic carries a fee and on which rate it is computed, and whether any of your traffic can be moved onto the vendor's own account without your setting it.

## Sources

1. Pricing — OpenRouter. https://openrouter.ai/pricing (accessed 2026-10-13)
2. BYOK — OpenRouter. https://openrouter.ai/docs/guides/overview/auth/byok (accessed 2026-10-13)
3. Stripe agrees to acquire OpenRouter to help businesses optimize token routing and usage — Stripe. https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter (accessed 2026-10-13)
4. AI Gateway Pricing — Vercel. https://vercel.com/docs/ai-gateway/pricing (accessed 2026-10-13)
5. AI Gateway pricing — Cloudflare. https://developers.cloudflare.com/ai-gateway/reference/pricing/ (accessed 2026-10-13)
6. Unified Billing — Cloudflare. https://developers.cloudflare.com/ai-gateway/features/unified-billing/ (accessed 2026-10-13)
7. Pricing — Requesty. https://www.requesty.ai/pricing (accessed 2026-10-13)
8. Plans and prices — Eden AI. https://www.edenai.co/docs/v3/overview/plans-prices (accessed 2026-10-13)
9. Pricing — Kong. https://konghq.com/pricing (accessed 2026-10-13)
10. Pricing — Portkey. https://portkey.ai/pricing (accessed 2026-10-13)
11. LiteLLM Enterprise — BerriAI. https://www.litellm.ai/enterprise (accessed 2026-10-13)
12. Pricing — Not Diamond. https://www.notdiamond.ai/pricing (accessed 2026-10-13)