LLM Gateway Comparison: LiteLLM, Portkey, OpenRouter and OriginCoder
Published 2026-11-11
An LLM gateway (also called an AI gateway) is a service between your application and model providers. It gives you one API for many models, and usually adds key management, usage tracking, routing and fallbacks. This guide compares four common options at the level of what kind of thing each one is, so you can narrow the choice before you read the details on each project's own site.
Disclosure: we build OriginCoder, so treat our description of it as the least neutral part. We only describe the others in general terms we are confident about. Features and pricing change often. Check each project's documentation before deciding. This comparison is current as of November 2026.
First, decide what you need
- One API for several providers. Do you want to call Claude, GPT and Gemini through a single request format?
- Who runs it. Do you want to operate the gateway yourself, or have someone else run it?
- Who pays whom. Do you want one invoice from the gateway, or to use your own provider accounts and keys?
- Controls. Do you need per-team keys, budgets, logs and audit trails?
- Coding agents. Will tools such as Claude Code or Codex connect to it?
The four options at a glance
| LiteLLM | Portkey | OpenRouter | OriginCoder | |
|---|---|---|---|---|
| What it is | Open-source SDK and proxy server you host | Gateway and control platform, with an open-source gateway and a hosted service | Hosted marketplace-style API for many models | Hosted gateway with its own balance and subscriptions |
| Who runs it | You | You (open-source) or Portkey (hosted) | OpenRouter | OriginCoder |
| Whose provider accounts | Yours, you bring provider keys | Yours, you bring provider keys | OpenRouter's, you buy credits (check their docs for bring-your-own-key options) | OriginCoder's, you pay in USD |
| Request format | OpenAI-style, many providers | OpenAI-style, many providers | OpenAI-style | OpenAI, Anthropic and Gemini formats |
| Strength | Full control and customisation | Observability, guardrails and governance features | Wide model choice behind one key | Simple billing and coding-agent setup |
| Main cost to you | Engineering and operations time | Platform fees if hosted | Per-token usage through credits | Per-token usage, or a subscription |
LiteLLM: self-hosted and flexible
LiteLLM is open source. You can use it as a Python library, or run its proxy server to give your team an OpenAI-compatible endpoint in front of many providers, with virtual keys, budgets and spend tracking. Because you run it, you control where your data goes and how it is configured. In exchange you operate it: deployment, a database for state, upgrades, monitoring and scaling are your job. It fits teams that already run infrastructure and need custom routing or strict data control.
Portkey: governance and observability
Portkey positions itself as an AI gateway and control plane. It offers an open-source gateway you can run yourself and a hosted platform that adds features such as logging, guardrails and prompt management. It fits teams that care about observability and policy across many providers and already hold their own provider accounts. Check Portkey's documentation for exactly which features are in the open-source gateway and which are hosted.
OpenRouter: many models behind one key
OpenRouter is a hosted service that exposes a large catalogue of models from many providers through one OpenAI-style API and one account. You buy credits and use them across models, and it can route among providers for a model. People searching for an "OpenRouter API" usually want exactly this: one key, many models, no separate provider signups. It fits experiments, side projects and apps that want to try many models fast.
OriginCoder: billing and coding-agent focus
OriginCoder is a hosted gateway. It accepts OpenAI, Anthropic and Gemini request formats on one endpoint, bills in USD pay-as-you-go or by subscription, records per-request usage with token counts and cost, and offers one-click import into CC-Switch for tools such as Claude Code and Codex. It fits developers who mainly want to run coding agents through one key and one balance. It is not a self-hosted product, and it does not replace the governance features of a platform built around them. See current per-model prices on the model pricing page and the setup in How to use Claude Code with a custom API endpoint.
How to choose
- You must keep everything on your own infrastructure: a self-hosted gateway such as LiteLLM, or the open-source Portkey gateway.
- You already have provider accounts and want logs, budgets and guardrails on top: a control plane such as LiteLLM or Portkey.
- You want many models behind one key with no setup: a hosted marketplace such as OpenRouter, or a hosted gateway such as OriginCoder.
- You mainly use Claude Code or Codex and want simple billing: a hosted gateway with coding-agent setup, such as OriginCoder. See also Claude Code alternative: official subscription vs. API gateway.
Whichever you pick, run a small real workload first and compare cost, latency and error rates on your own traffic. Ask any provider where your prompts are stored and for how long.