What every major LLM gateway actually charges in 2026 — the credit-purchase fees, per-token markups, subscription costs, EU data residency, and who accepts crypto. Numbers only, no vendor marketing.
· llmdeal.me
The LLM gateway space looks crowded until you start reading pricing pages carefully. Fees are structured very differently across providers — some charge at credit purchase, some bake a markup into every token, some charge a SaaS subscription and nothing per token. Depending on your usage pattern, picking the wrong structure can cost you 5–10% more than the alternative for identical API calls.
This post covers the nine providers that come up most in developer discussions: OpenRouter, Requesty, Portkey, Helicone, LiteLLM, Cloudflare AI Gateway, Martian, Vercel AI Gateway, and Kong AI Gateway. For each one, the focus is on the business model and what you actually pay — not feature counts.
Fee structures at a glance. "Markup" means an extra charge on top of base provider token prices. "Gateway fee" means a per-request or per-log charge separate from tokens. "Subscription" means a flat monthly cost.
| Provider | Model | Fee / markup | EU residency | Crypto |
|---|---|---|---|---|
| OpenRouter | Credit marketplace | 5.5% on card purchase / 5% on USDC; zero per-token markup | Enterprise only (opt-in) | USDC |
| Requesty | Per-token markup | 5% on every token (e.g. $10/M → $10.50/M) | Frankfurt (AWS eu-central-1) — all EU traffic | No |
| Portkey | Observability SaaS | Zero per-token; $49/mo (Production), enterprise from ~$2k/mo | Enterprise tier only | No |
| Helicone | Observability SaaS | Zero per-token; free → $79/mo (Pro) → $799/mo (Team) | Not advertised | No |
| LiteLLM (self-host) | Open-source proxy | $0 license; real TCO ~$2k–$3.5k/mo (infra + devops) | Fully controlled (you choose) | N/A |
| Cloudflare AI Gateway | Infrastructure layer | Zero per-token; free (100k logs/mo) or $5/mo (10M logs/mo) | Via Regional Services (separate product) | No |
| Martian | AI router | ~5.5% passthrough fee (third-party estimate); enterprise custom | Not found | No |
| Vercel AI Gateway | Platform add-on | Zero per-token; requires Vercel Hobby/Pro ($20/mo) or Enterprise | Not offered | No |
| Kong AI Gateway | Enterprise API mgmt | ~$105/mo per LLM service (Konnect SaaS); enterprise from $30k/yr | Not specified | No |
OpenRouter is the dominant multi-model marketplace with 400+ models and 25+ available on a free tier. Its pricing structure is deliberately front-loaded: the fee is charged when you buy credits, not per token.
On EU data residency: standard-tier requests run on Cloudflare's global edge with no residency guarantee. EU-resident processing is available only to enterprise customers via a dedicated opt-in. One practical consequence: teams that need GDPR-compliant data routing cannot use OpenRouter's standard tier and document compliance in an audit trail.
One minor note: a third-party site (costbench.com) claimed OpenRouter applies a 100% markup to Claude models. OpenRouter's own FAQ and all other sources state zero per-token markup, so that claim is almost certainly incorrect.
Requesty is the only major provider among this group that charges a flat percentage markup baked into the token price itself, rather than at credit purchase or as a subscription.
Stated on their pricing page verbatim: "5% markup on base model costs. No subscription, no seat fees, no minimum spend." So if a model costs $10/M tokens at the provider, you pay $10.50/M through Requesty. Over 400 models, smart routing by request type (code, summarization, reasoning), and claimed cost reductions of up to 40% via semantic caching.
The EU hosting story is the most concrete of any provider here: all EU requests are exclusively processed in Frankfurt, Germany on AWS eu-central-1, reachable at router.eu.requesty.ai. SOC 2 Type II certified, DPA available on request. They raised $3M in 2024 and report over 1,200 EU customers.
No crypto payment option was found. Requesty's pricing is also straightforward to evaluate: if your monthly token spend is X, your Requesty cost is X × 1.05. Compare that to OpenRouter's 5.5% paid up-front at credit purchase with zero per-token overhead.
These two are in a different category: their core value is logging, tracing, and analytics around your LLM calls, not model access itself. You connect them to providers via your own API keys; the gateway layer adds zero per-token cost.
Portkey earns revenue on its observability subscription. The Developer plan is free at 10,000 logged requests/month with 3-day retention. Production is ~$49/month base plus $9 per additional 100k logs. Enterprise tiers — which include SOC2 Type 2, ISO 27001, GDPR, HIPAA, and EU data residency — run from roughly $2,000 to $10,000+/month by one 2026 estimate. Portkey was named a Gartner Cool Vendor in LLM Observability in 2025 and supports 1,600+ models via its open-source gateway. GDPR and EU hosting are only enforced at Enterprise.
Helicone (YC W23, open-source) has a similar structure but a cleaner published pricing ladder: free (10k requests/month, 7-day retention), Pro at $79/month (unlimited seats, alerts, 1-month retention), Team at $799/month (5 orgs, SOC-2, HIPAA, 3-month retention), and Enterprise custom with on-premises options. Zero per-token markup at all tiers. EU data residency is not advertised. Startups get 50% off the first year; students and open-source maintainers can apply for free access.
If you are already managing API keys directly with providers and only need observability, either of these is a reasonable pick — you add zero marginal cost per token and get a structured log layer. The trade-off is that compliance features (EU hosting, HIPAA) live behind expensive enterprise plans.
LiteLLM is MIT-licensed and self-hosted. The license fee is $0. What is not zero is the operational cost. A 2026 analysis by Truefoundry put the realistic total cost of ownership for a production LiteLLM deployment at $2,000–$3,500/month when you account for cloud infrastructure, devops time, monitoring, and incident response.
If you want someone else to run it, there are two published enterprise tiers — though pricing is not listed on the official docs page and requires contacting [email protected]. Third-party analysis puts Enterprise Basic at approximately $250/month (adds SSO, guardrails, Prometheus metrics, JWT auth) and Enterprise Premium at approximately $30,000/year (~$2,500/month) with priority support and a dedicated account manager.
For EU data residency: because you self-host, you control where the proxy runs entirely. Point it at a Frankfurt VM and your processing stays in Frankfurt. This is the cleanest GDPR story of any option here — at the cost of owning all operational complexity.
Cloudflare AI Gateway charges nothing for the gateway itself and nothing per token. The cost is in log storage: 100,000 total AI Gateway logs/month on the free Workers tier, 10,000,000 logs per gateway on Workers Paid at $5/month. Overages on paid plans are $0.05 per million additional requests.
Twenty-plus AI providers are supported. The routing logic is rule-based (segment, geography, A/B weighting) rather than AI-driven cost optimisation. EU data residency is technically possible, but only via Cloudflare's separate Regional Services / Data Localization Suite product — it is not automatic or included in the AI Gateway feature itself.
This is a strong option if you are already deep in the Cloudflare ecosystem and want a lightweight proxy layer. It is not a substitute for OpenRouter-style model discovery or Requesty-style smart routing.
Three more worth knowing, each with a narrow target audience.
Martian is a pure AI router: it analyses each prompt in real-time and sends it to the model it thinks will handle it best. Cost reduction claims range from 20% to 97%. Their valuation was reportedly approaching $1.3B in April 2026, which is a large number for a routing layer. The platform fee is approximately 5.5% based on a third-party comparison — Martian's own pricing page was unreachable during research, so treat that figure as an estimate. Enterprise pricing is opaque: custom contracts, VPC deployment options, SLA. No EU residency guarantee found, no crypto payment.
Vercel AI Gateway exists for Vercel platform users. Zero per-token markup, zero BYOK fee, $5/month in free credits (stops once you make a first payment). But it requires a Vercel plan: Hobby (free), Pro ($20/month), or Enterprise. It is not a standalone product and has no EU data residency offering.
Kong AI Gateway is squarely enterprise API management with an LLM add-on. At roughly $105/month per Gateway Service on Konnect SaaS — where each LLM provider counts as one service — and enterprise contracts starting at $30,000–$50,000/year, Kong is targeting teams that already run Kong for their API infrastructure, not developers looking for a cheap LLM access layer.
Among all of the above, only two providers publicly accept crypto payment: OpenRouter (USDC only, via on-chain x402 payments, non-refundable) and AIMLAPI (Bitcoin plus 300+ additional coins). No gateway among the major players explicitly advertises a no-KYC signup flow.
For EU-based developers who want both EU data residency and the ability to purchase credits without a credit card or identity verification, the supply is thin. Requesty has the strongest EU residency story of the lot; it does not accept crypto. OpenRouter accepts USDC but routes standard-tier traffic on a global edge without residency guarantees. The intersection — EU-resident processing combined with crypto/no-KYC payment — is a near-empty spot in the market.
One regulatory note: US Bureau of Industry and Security (BIS) rulemaking on IaaS providers may eventually require identity verification from AI infrastructure and API platforms for foreign customers. How and whether that applies to LLM gateways has not been tested, but it is worth tracking if no-KYC access is a requirement for your stack.
llmdeal.me is designed specifically for that intersection: EU-resident processing, OpenAI-compatible single endpoint, crypto checkout, no KYC.
Run through these questions in order:
The fee structures look similar at the top level — 5–5.5% is where several land — but the mechanics differ enough to matter. OpenRouter's fee is paid when you buy credits (front-loaded, one-time). Requesty's fee accumulates on every token every call (ongoing, compounding as usage scales). Subscription-based tools like Portkey and Helicone are volume-agnostic once you pay the monthly rate. And self-hosted LiteLLM trades money for engineering time, which only makes economic sense past a certain scale.
Rates checked against providers' own pricing pages, May 2026. Article published 2026-05-16.