2027 planning guide — evidence checked 20 November 2026

Best AI Gateway for 2027: Routing, Cost and Control

A gateway sits between your application and one or more model providers. This guide compares the fees, routing controls, observability and exit paths of seven options, including the case for not using a gateway at all.

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.

AI gateways and LLM routers for 2027 planning, checked 20 November 2026
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.

Monthly gateway cost at three inference-spend levels, calculated 20 November 2026
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

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.

See the models Load balancing

Related reading: unlimited API access, spend controls, AI API pricing questions, compare Plugsky.