Short answer. Add an AI gateway when you need multi-provider routing, fallback, caching or per-team spend controls. OpenRouter leads on model breadth, Cloudflare and Vercel on low-friction pass-through, Kong on API governance and LiteLLM on self-hosting. Plugsky is not a gateway; it is one upstream API that can remove the need for one.
A gateway is infrastructure, not a model. It adds a hop, a control plane and a bill. In exchange, it can route around a failed provider, cache repeated prompts, enforce budgets and give finance one place to see spend. The value depends on how many providers you actually use and how much control you need.
This is a 2027 planning guide with evidence checked on 20 November 2026. It does not claim 2027 tests or awards. It separates the gateway fee from the inference fee, because a "free" gateway can still carry a credit-purchase surcharge or a per-model charge.
How we evaluated this
We weighted five criteria for a team that already calls at least two model providers, or expects to.
- Routing control, 25%. Fallback, weighted routing, model aliases, provider allowlists and the ability to pin traffic to a region.
- Fee transparency, 20%. Whether the gateway fee is separate from inference, and whether the published rate card is enough to budget.
- Observability, 20%. Request logs, token and cost analytics, export, retention and alerting.
- Exit and self-hosting, 20%. Whether the buyer can run the gateway, export configuration and avoid a proprietary request format.
- Model coverage, 15%. The number and type of providers and models reachable through one endpoint.
We did not test failover under a real provider outage, measure added latency, or verify cache hit rates. We did not review enterprise contracts, SSO configurations or log-retention terms. We did not confirm that every listed model is available in every region. The comparison uses published terms, not production telemetry.
Seven gateway and routing options compared
"Inference markup" is the amount added to the provider's token price. "Gateway fee" is what the gateway charges separately. A dash means the provider does not publish a comparable figure.
| Option | Role | Gateway fee | Inference markup | Routing and fallback | Visibility / controls | Self-host | Best fit |
|---|---|---|---|---|---|---|---|
| OpenRouter | Multi-provider marketplace and router | 5.5% on Standard credit purchases; 8% on Business | None; inference at provider list price | Auto-routing, preferred providers, fallback, budgets | Activity logs and spend controls | No | Broad model access with one bill |
| Cloudflare AI Gateway | Gateway in front of provider APIs | Core features free; 5% fee on Unified Billing credits | None; provider rates passed through | Dynamic routes, provider allowlist, rate limiting | Analytics, logs, caching, DLP | No | Cloudflare estates and free core controls |
| Vercel AI Gateway | Gateway for Vercel and external apps | 0% markup on paid tier; 5% regional endpoint surcharge | None on paid tier | Provider routing and fallback; provider allowlist | Reporting, tags, spend controls | No | Vercel and Next.js applications |
| Kong AI Gateway | API management plus AI gateway | AI Management Plus from $25/month plus usage; $100/month per model beyond 5 | None stated; inference billed by provider | AI proxy, rate limiting, semantic caching, MCP/A2A governance | Analytics, logs, policy controls | Gateway Enterprise self-hosted by quote | Regulated API governance |
| LiteLLM | Open-source gateway and proxy | Free to self-host; Enterprise priced by request capacity | None | Routing, fallback, budgets, model aliases | Logs, spend tracking, audit logs on Enterprise | Yes | Self-hosted and air-gapped control |
| Portkey / Palo Alto | AI gateway now under Palo Alto Prisma AIRS | Not published on the redirected page | Not published on the redirected page | Enterprise AI control plane; confirm current standalone scope | Enterprise security and governance features | Not stated on the redirected page | Palo Alto security estate |
| Plugsky | Single upstream API, not a gateway | No separate gateway fee | None; flat plan with no monthly token cap | Documented model fallback chains inside the API | Provider-side usage; gateway features depend on your client | Private deployment by contract | Removing the need for a gateway |
Compare the total, not the headline. OpenRouter charges no inference markup but takes 5.5% when you buy Standard credits. Cloudflare's gateway is free but applies a 5% fee to Unified Billing credits. Vercel charges no markup on its paid tier but adds 5% for a regional endpoint. Kong's entry price is low, but each model beyond five costs $100 per month. The cheapest gateway depends on spend and model count, not the logo.
Cost math: $3,000, $10,000 and $50,000 in monthly inference
Assume the buyer spends the stated amount on inference at provider list price, uses eight models, and buys credits through the gateway where that is the published mechanism. The table isolates gateway cost and excludes infrastructure, support and taxes.
| Option | $3,000 spend | $10,000 spend | $50,000 spend |
|---|---|---|---|
| OpenRouter Standard (5.5%) | $165 | $550 | $2,750 |
| OpenRouter Business (8%) | $240 | $800 | $4,000 |
| Cloudflare Unified Billing (5%) | $150 | $500 | $2,500 |
| Vercel regional endpoint (5%) | $150 | $500 | $2,500 |
| Kong Plus, 8 models | $325 | $325 | $325 |
| LiteLLM open source | $0 licence | $0 licence | $0 licence |
| Plugsky Builder | $60 | $60 | $60 |
At $3,000 of spend, OpenRouter Standard costs $165 and Cloudflare or Vercel cost $150. At $50,000, the percentage fees dominate: $2,750, $2,500 or $2,500 per month. Kong's $325 is flat in this model because the model count is fixed, but Kong also charges usage, which is not included. LiteLLM's licence is free, but the buyer pays for compute, staff and support.
Plugsky is not a like-for-like row. It replaces the upstream provider spend with a flat plan rather than routing to other providers. At $3,000 of equivalent inference, Builder is $60, but it cannot route to OpenAI or Anthropic. The comparison shows the value of consolidation, not a gateway discount.
When a competitor is the better choice
OpenRouter is the better choice when model breadth is the priority: 600+ models, 80+ providers, auto-routing and one bill. It is also the better choice when the team wants to compare providers without signing each one. Cloudflare AI Gateway wins when the application already runs on Cloudflare and the team wants free caching, analytics and rate limiting with no inference markup. Vercel AI Gateway wins for Vercel and Next.js teams that want zero markup on the paid tier and tight integration with deployments.
Kong wins when the gateway must sit inside an existing API-management programme with policies, developer portals and enterprise governance. LiteLLM wins when self-hosting, air-gapping or an open-source exit path is mandatory. Portkey, now under Palo Alto Networks Prisma AIRS, may fit buyers already standardized on Palo Alto security; standalone pricing was not published on the page checked, so a quote is required.
What we did not verify
We did not simulate a provider outage or measure failover time. We did not measure the latency a gateway adds to streaming responses. We did not verify cache hit rates, log retention, export formats or alerting. We did not review SSO, SCIM, RBAC or audit-log behaviour. We did not confirm the number of models actually reachable in each region, or whether every model supports tools, JSON mode and vision.
For Plugsky, the API has documented fallback chains, but that is provider-side routing, not a multi-provider gateway. A SOC 2 Type II report is available under NDA; confirm the current scope and audit period. Request the current report under NDA. HIPAA/BAA is available by contract and subject to review. The 99.9% figure is an uptime target, and the contract schedule governs.
Frequently asked questions
Do I need an AI gateway?
Only if you use multiple providers, need fallback or caching, or must enforce budgets across teams. A single-provider application can call the provider directly and avoid an extra hop, bill and failure point.
Which AI gateway has the lowest fee?
LiteLLM is free to self-host, but you pay for infrastructure and operations. Cloudflare's core gateway is free with a 5% credit-purchase fee, and Vercel charges no markup on its paid tier. Kong starts at $25 per month plus usage.
Does OpenRouter mark up model prices?
No. OpenRouter states that inference is billed at the provider's list price. It charges 5.5% when you buy Standard credits and 8% on Business, plus a 5% fee above the BYOK allowance.
Can I self-host an AI gateway?
Yes. LiteLLM is open source and free to self-host, with an Enterprise tier priced by request capacity. Kong offers a self-hosted Gateway Enterprise option by quote. Cloudflare, Vercel and OpenRouter are managed services.
Is Plugsky an AI gateway?
No. Plugsky is a single OpenAI-compatible API with 39 model IDs. It can reduce the number of upstream providers you need, but it does not route to OpenAI, Anthropic or Google. If you need multi-provider routing, use a gateway in front of it.
What should I measure before choosing a gateway?
Measure added latency, failover time, cache hit rate, log retention, cost per request and the engineering time to operate it. Also confirm the fee basis: percentage of spend, per request, per model or per seat.
Does a gateway improve reliability?
It can, if it has health checks and a tested fallback path. A gateway without tested fallback adds a failure point. Ask for the failover behaviour during a provider outage and test it before production.
Sources
- OpenRouter pricing — checked 20 November 2026.
- Cloudflare AI Gateway pricing — checked 20 November 2026.
- Vercel AI Gateway pricing — checked 20 November 2026.
- Kong pricing and AI Management plans — checked 20 November 2026.
- LiteLLM pricing — checked 20 November 2026.
Remove a hop from the stack
If your only reason for a gateway is to reach several model classes, one API with 39 model IDs may be simpler. Check the catalogue and the fair-use limits.
Related reading: unlimited API access, spend controls, AI API pricing questions, compare Plugsky.