OpenRouter Alternatives: Pick by Why You're Leaving
September 30, 2026
Most lists of OpenRouter alternatives rank a dozen tools side by side. That's the wrong way to choose, because the tools on those lists solve different problems. Pick an alternative by the reason you're leaving.
There are four common reasons, and each one points to a different kind of tool:
- You want the gateway on your own servers. Self-host an open-source proxy like LiteLLM.
- You want routing and logs, but with your own provider accounts. Put a managed gateway like Cloudflare AI Gateway or Vercel AI Gateway in front of your keys.
- Your company wants one vendor contract. Go direct to a provider, or through Amazon Bedrock, Google's Gemini Enterprise Agent Platform (formerly Vertex AI), or Microsoft Foundry.
- You want no provider at all. Run open models locally.
First, be clear about what you're replacing.
What OpenRouter actually does
OpenRouter is a hosted proxy. You send requests in the OpenAI format to one base URL, and it forwards them to the model provider. You can point the OpenAI SDK at it and change nothing else.
| Piece | How OpenRouter handles it (September 2026) |
|---|---|
| API | OpenAI-compatible, one base URL for hundreds of models |
| Provider routing | By default it load-balances across providers, favouring stable and cheaper ones. You can set an order, sort by price, throughput or latency, or turn fallbacks off |
| Model fallbacks | Pass a list of models. If the first errors, it tries the next, and you pay for the model that answered |
| Billing | Prepaid credits. Inference is passed through at the provider's price with no markup. The fee sits on buying credits |
| BYOK | Bring your own provider keys. Your key is tried first. A service fee applies once usage passes a monthly allowance |
| Privacy defaults | Prompts and completions aren't logged unless you opt in. A zdr flag limits routing to endpoints that don't retain prompts |
That last row matters more than people expect. The provider routing docs let you deny providers that store data, per request. A lot of "I'm leaving for privacy" decisions are really "I didn't know that setting existed."
So OpenRouter bundles three things: one API, a routing layer, and one bill. Each alternative below keeps some of those and drops others.
OpenRouter alternatives: quick comparison
| Shape | Who runs it | Whose provider accounts | Best reason to pick it | |
|---|---|---|---|---|
| OpenRouter | Hosted proxy + marketplace | OpenRouter | Theirs, or yours with BYOK | Fast access to many models on one bill |
| LiteLLM | Open-source proxy and Python SDK | You | Yours | Data never leaves your infrastructure except to the provider |
| Portkey gateway | Open-source gateway (MIT) | You, or Palo Alto Networks as Prisma AIRS AI Gateway | Yours | Guardrails and governance in the gateway |
| Cloudflare AI Gateway | Managed gateway at the edge | Cloudflare | Yours, or Cloudflare's unified billing | Caching, rate limits, logs in front of existing apps |
| Vercel AI Gateway | Managed gateway | Vercel | Vercel's, or yours with BYOK | Budgets, failover and logs without running a proxy |
| Bedrock / Gemini Enterprise Agent Platform / Microsoft Foundry | Cloud model platform | Your cloud provider | Your cloud account | One contract, one compliance boundary |
| Local runtimes | Model server on your hardware | You | None | No per-token bill and no third party |
If you want control: self-host the gateway
A hosted proxy means every prompt passes through one more company. Fine for a side project. For a product that handles customer data, legal or security may want the router inside your own network.
LiteLLM is the closest self-hosted replacement. It comes in two forms: a Python SDK you call from code, and a proxy server you run centrally. Both speak the OpenAI format and translate to OpenAI, Anthropic, Gemini, Bedrock, Azure and many more.
The proxy is the part that replaces OpenRouter. It gives you:
- Virtual keys, so each team or app gets its own key without seeing the real provider key
- Spend tracking and budgets per project
- Load balancing, retries and fallbacks across providers and deployments
- Logging integrations to send traces wherever you already look
The core is MIT-licensed; enterprise features sit under a separate commercial licence, so check which ones you need. The trade-off: you now run production infrastructure. Upgrades, uptime and the database behind spend tracking are yours.
Portkey is the other open-source option people reach for. Its gateway handles fallbacks, retries and load balancing, and adds guardrails for checking inputs and outputs. One change to know about: Palo Alto Networks completed its acquisition of Portkey on May 29, 2026, and it's now the AI gateway inside Prisma AIRS. The open-source repo is still MIT-licensed.
| LiteLLM | Portkey gateway | |
|---|---|---|
| Licence | MIT core, commercial enterprise features | MIT |
| Also ships as | A Python SDK you call from code | Part of Prisma AIRS (Palo Alto Networks) |
| Standout | Virtual keys and per-project budgets | Guardrails on inputs and outputs |
| Deploys on | Docker, with Terraform modules for AWS and GCP | Docker, Node.js, Kubernetes, major clouds |
If you want a gateway in front of your own keys
Maybe a managed service is fine. You just want to keep your own provider accounts and negotiated rates, and still get logs, caching and fallbacks.
Cloudflare AI Gateway
Cloudflare AI Gateway sits between your app and the provider. It's available on all Cloudflare plans. You get analytics on requests, tokens and cost, logging, caching, rate limiting, and request retry with model fallback.
It has an OpenAI-compatible chat completions endpoint that works across providers including OpenAI, Anthropic and Google. You can authenticate three ways: bring your own provider keys, pass provider keys per request, or use Cloudflare's Unified Billing and pay Cloudflare for tokens. Cloudflare says its core gateway features are free. Paid parts are log storage beyond your plan's limit and a fee on credits bought through Unified Billing.
Caching is the stand-out. If your app sends the same request repeatedly, Cloudflare can serve it from its own cache instead of calling the provider again. That's faster, and you don't pay for the repeat.
Vercel AI Gateway
Vercel AI Gateway is the closest managed match to OpenRouter's shape: one API key, many providers, a credit balance. It speaks the OpenAI Chat Completions and Responses formats, the Anthropic Messages format, and Vercel's own AI SDK.
Four details worth getting right:
- You don't need to host on Vercel. The docs say you can call it from any server, cloud or local environment with an API key.
- Vercel says it adds zero markup to provider token prices, including with BYOK. You pay list price from prepaid credits.
- Budgets can be set per team, project, API key or member. They only cover spend on Vercel's credentials, not your BYOK spend.
- It logs every routing attempt with provider, latency, tokens and cost, and supports ordered provider and model fallbacks.
| Cloudflare AI Gateway | Vercel AI Gateway | |
|---|---|---|
| Main strength | Caching and rate limiting at the edge | Budgets and failover on one credit balance |
| Your own keys | Yes | Yes (paid tier) |
| Pay the gateway for tokens | Optional, via Unified Billing | Default, via credits |
| Must host with them | No | No |
If you want one vendor contract
Sometimes the reason is procurement. Your company already has a cloud agreement, a data-processing addendum and a security review. Adding a new AI vendor means doing all of that again. In that case, the alternative isn't another gateway. It's the model platform on the cloud you already use.
| Platform | Model range (September 2026) | API shape |
|---|---|---|
| Amazon Bedrock | Amazon, Anthropic, DeepSeek, Moonshot AI, MiniMax, OpenAI, xAI and more | Converse API across models; the docs also show OpenAI SDK calls (Chat Completions, Responses) on the runtime endpoint |
| Gemini Enterprise Agent Platform (formerly Vertex AI) | Google models, partner models including Claude, and managed open models in Model Garden | Native APIs, plus OpenAI-library access to Gemini models |
| Microsoft Foundry (formerly Azure AI Foundry) | Microsoft, OpenAI, Anthropic, Meta and others | One project endpoint using the OpenAI client |
Two catches. Model lists differ by platform and region, so confirm the exact model is served where your data has to live. And you lose cross-provider failover unless you add a gateway on top, such as LiteLLM in front of Bedrock and Foundry.
Going direct to a single provider is the simplest version of this. If one model family covers all your needs, a gateway may be a layer you don't need.
If you want no provider at all: run models locally
The last reason is cost or privacy at the extreme. You don't want any per-token bill, or no prompt may leave the machine. Then you want a local runtime, not a gateway.
The good news: most local runtimes expose an OpenAI-compatible endpoint too, so code written for OpenRouter often needs only a base-URL change. The choice is between a desktop-friendly runner and a serving engine built for many users. Our vLLM vs Ollama comparison covers that split, and our Ollama alternatives guide covers the desktop side.
The catch is hardware. Frontier-sized open models don't fit on a laptop, and closed models like Claude or GPT don't run locally at all. A mix is often the practical answer: a local model for private work, a gateway for the rest.
How to decide
| Your real reason | Start here |
|---|---|
| "Legal won't let prompts pass through a third-party proxy" | LiteLLM, self-hosted |
| "We need guardrails and governance in the gateway" | Portkey gateway |
| "We want caching and rate limits in front of our own keys" | Cloudflare AI Gateway |
| "We like OpenRouter's shape but want per-project budgets" | Vercel AI Gateway |
| "Procurement wants one cloud contract" | Bedrock, Gemini Enterprise Agent Platform or Microsoft Foundry |
| "We want no per-token bill" | A local runtime |
If none of these match, stay. OpenRouter is still the fastest way to try many models on one bill, and its routing and privacy settings cover more than most people use. For tracing and evals, a gateway's logs are a start, but a dedicated tool goes further. See Langfuse vs LangSmith.
Where Taku fits
Gateways decide which model answers a request. They don't help with the step before that: knowing which setup is worth running. Taku is an AI-native desktop workspace where you mirror an AI app, skill or workflow someone already got working, run it, remix it and keep it. It's not a gateway and doesn't replace one. Browse the free apps to see the shape. Taku is in Beta, and the Mac app is available now.
FAQ
What is the best self-hosted OpenRouter alternative?
LiteLLM's proxy is the closest match. It gives you one OpenAI-compatible endpoint, virtual keys, budgets and fallbacks on your own infrastructure. Portkey's MIT-licensed gateway is the main alternative if guardrails matter more to you.
Is OpenRouter more expensive than going direct?
Not on tokens. OpenRouter says it passes provider pricing through with no markup on inference. The cost is a fee when you buy credits, and a fee on BYOK usage past a monthly allowance. Check its FAQ for current terms.
Does Vercel AI Gateway only work on Vercel?
No. Vercel's docs say you can call it from any server, cloud or local environment with an AI Gateway API key. Apps deployed on Vercel can also authenticate with OIDC.
What is an LLM gateway?
A service between your app and model providers. It gives you one API, and usually adds routing, fallbacks, logging, caching and spend controls. OpenRouter, LiteLLM, Portkey, Cloudflare AI Gateway and Vercel AI Gateway are all LLM gateways with different owners and billing models.
Is Vertex AI still called Vertex AI?
Google renamed the Vertex AI Platform to Gemini Enterprise Agent Platform. Model Garden is still there under the new name, with Google, partner and open models.
Can I switch away from OpenRouter without rewriting my code?
Usually yes. OpenRouter, LiteLLM, Cloudflare's compatible endpoint, Vercel AI Gateway and most local runtimes accept the OpenAI format. Changing the base URL, the key and the model names is often enough. Test streaming and tool calls before you move production traffic.