Which AI gateway should you pick in 2026?

Last updated 23 August 2026. I opened Vercel, OpenRouter, LiteLLM, and Portkey docs the same day I rewrote this. Model counts come from public APIs I called, not from memory. Ellelion runs on Vercel AI Gateway, so the bias is in the open.

What question is this answering?

I have picked an AI provider for every product I ship, and I have changed my mind more than once. This post is the version with the sales deck peeled off: what each option reached on 23 August 2026, what it charged, and where I land after building on them. If you want the hosting story, read Why I run the Ellelion family on Vercel first. The company index is ellelion.com.

What is the difference between the AI SDK, a gateway, and a provider?

People dump these names in one bucket. They are three jobs. The AI SDK is a TypeScript library. It is how your code calls a model, streams tokens, and runs tools. A gateway such as Vercel AI Gateway or OpenRouter sits between that code and the model vendors. You get one endpoint, one bill, and failover. A provider such as Groq or Together is who runs the weights. Groq publishes LPU speed claims on groq.com. Those are Groq's numbers. I have not reproduced them here.

You can stack all three: SDK in the repo, gateway for routing, provider underneath. Sort them apart and the rest of this gets easier.

What does Vercel AI Gateway cost?

This is what Ellelion runs on. Vercel marked AI Gateway generally available in August 2025. On 23 August 2026 I called https://ai-gateway.vercel.sh/v1/models and counted 352 models in the data array. Older writeups said 200+. That number is dead. Use the list.

The pricing page, opened the same day, is blunt:

AI Gateway charges no markup and no platform fee on tokens.Vercel Docs, AI Gateway Pricing

Source: Vercel Docs, "AI Gateway Pricing" (opened 23 August 2026), https://vercel.com/docs/ai-gateway/pricing

BYOK is listed the same way: no markup and no platform fee on tokens. You still pay Gateway request charges and any add-on you enable. Routing stays simple on purpose. You get provider failover and cost-aware model picks. You do not get weighted load-balancing across ten keys or per-request budget policies. For the way I ship, that ceiling has been high enough. A platform team routing millions of calls with strict cost rules would feel it.

What does OpenRouter charge beyond tokens?

OpenRouter is the one to beat on raw breadth. On 23 August 2026 I called https://openrouter.ai/api/v1/models and the response reported total_count 422.

Their FAQ, opened the same day, says:

We pass through the pricing of the underlying providers; there is no markup on inference pricing (however we do charge a fee when purchasing credits).OpenRouter FAQ, Pricing and fees

Source: OpenRouter FAQ, "Pricing and fees" (opened 23 August 2026), https://openrouter.ai/docs/faq

The same FAQ says BYOK has a plan-dependent free allowance measured by list-price inference cost, not request count. The HTML they ship strips the exact percentages in the static markup, so I am not going to invent 5.5% from an older draft. If you need the current credit-purchase fee, open their FAQ and read the live numbers before you commit.

When should you self-host LiteLLM or Portkey?

Managed gateways stop making sense the moment compliance or internal networking takes the wheel. LiteLLM is an open-source proxy that speaks the OpenAI format to many providers and runs on your own boxes. I am not repeating a GitHub star count here because I did not freeze a star snapshot for this rewrite. Portkey aims at teams that want observability, guardrails, and the option to self-host. You give up the zero-setup convenience. You get control over where traffic goes and who can see it. If a security review is going to ask where your prompts travel, this is the lane.

How do the four compare?

How do I choose for Ellelion's work?

The rule I use across twelve products is boring. If I am shipping on Vercel and writing TypeScript, I reach for the AI SDK and let the Gateway route, because the setup cost is near zero and the token bill matches provider pricing. If Vercel did not carry a model I needed, I would point the same AI SDK code at OpenRouter and keep moving. If a client needed prompts to never leave their network, I would stand up LiteLLM. None of these is a religion. The AI SDK is provider-agnostic on purpose, so swapping the gateway underneath is a config change.

Live Ellelion products that already speak MCP are Skillselion and Launchelion. The Agent Plugins Directory and botteams.io are live in the browser without a public MCP server yet.