AI gateway comparison

Choose a hosted gateway such as OpenRouter or Portkey if you want someone else to run it, and LiteLLM if you want a proxy you host yourself. Choose RelayPlane if what you need is a hard dollar cap on each agent run, enforced on your own machine.

We make RelayPlane, a free and open source proxy, so we say where the other options fit better.

How we compared them

An AI gateway is a service or proxy that sits between your code and the model providers. We judged six of them on four things a developer choosing one cares about. The options below are listed alphabetically, not ranked.

  • Where it runs and who operates it.
  • What the tool itself costs, as its own pages state it.
  • What it does about spend on a single agent run.
  • How hard it is to leave once your code points at it.

Everything said about another company comes from its own pages, fetched on 3 November 2026. Opinions are marked as ours.

Cloudflare AI Gateway

Cloudflare describes its gateway as a way to “gain visibility and control over your AI apps”, and says it is “Available on all plans”. The docs list caching, rate limiting, request retries and model fallback. It fits a team that already runs on Cloudflare and wants a gateway without a new vendor.

The drawback is on the spend side. The overview describes rate limiting as “limiting the number of requests your application receives”, which counts requests and does not say anything about dollars. In our view, check the full docs before you rely on it to stop one expensive run. We compare it in more detail on our Cloudflare AI Gateway comparison.

Helicone

Helicone calls itself “an AI Gateway & LLM Observability Platform for AI Engineers”. It is strongest when you want to see requests, sessions and users, with the gateway as part of that. Its pricing page lists a Hobby plan with “10,000 free requests” and a Pro plan at “$79 per month” marked “Usage-based pricing applies”.

The same page opens with the line “Helicone is no longer part of Mintlify. Details to come soon.” If you pick Helicone, read the current terms first. Our Helicone comparison covers the split between logging after the fact and refusing a request before it goes out.

LiteLLM

LiteLLM's docs describe “a single, unified interface to call 100+ LLMs” in the OpenAI format, plus a “Self-hosted LLM Gateway (Proxy) with virtual keys, cost tracking, and an admin UI”. It fits a team that wants to host its own gateway, serve many internal users and give each one a key.

In our view the cost is operational. You run it, patch it and size it yourself, and the full gateway with the admin UI is a bigger thing to stand up than a single local process. Our LiteLLM vs Portkey page and the LiteLLM comparison go further.

OpenRouter

OpenRouter's FAQ says it “passes through the pricing of the underlying providers” and that you add credits and pay from them. Its pricing page shows a platform fee of “Standard 5.5% Business 8%”, a free plan limited to “50 requests/day”, and “Budgets & Spend Controls” marked No on Free and Yes on Standard.

It fits a team that wants one bill and one API across many models with nothing to host. The trade is that every request goes through their service and you pay the platform fee on top of model prices. See our OpenRouter comparison and OpenRouter alternatives.

Portkey

Portkey's open source repo says “Route to 250+ LLMs with 1 fast & friendly API” and offers a hosted gateway. Its pricing page lists a free Developer plan with “10k recorded logs per month” and a Production plan at “$49/month”. The Developer plan is labelled “Not suitable for production workloads.”

The enterprise tier lists “Granular Budget & Rate Limits” under governance, so spend controls look aimed at larger organizations with contracts. It fits a team that wants guardrails, prompt management and logs in one hosted product. We compare it on our Portkey page and in Portkey alternatives.

RelayPlane

RelayPlane is a free, MIT licensed proxy that runs on your own machine. It meters what each agent run costs and blocks the request that would cross a cap. It is the right fit for developers running Claude Code, Cursor or their own agents who want a hard limit, and a poor fit if you need a hosted console, team seats or enterprise controls. These two lines come from the quickstart docs and the budget cap docs.

npm install -g @relayplane/proxy
relayplane start
relayplane cap set --day 10

A single run can carry its own limit with a per-run cap header, described in the runs docs. The cap is checked before a request is forwarded. Because the proxy is local and open source, leaving is a base URL change. The detail is on the LLM proxy page.

Which one fits whom

If you want nothing to operate, the hosted ones win. If you want one place to look at request logs, Helicone leads with that. If your stack already sits on Cloudflare, its gateway is the shortest path. If your worry is one agent loop burning the budget overnight, start with a local cap and add a hosted gateway later if you need one.

Sources

Prices change. Check each vendor's page before you decide.